最新刊期
-
Lightweight wide-range license plate recognition network with parallel encoder-decoder architecture AI导读
摘要：ObjectiveIn complex open scenarios with diverse license plate layouts， backgrounds and character lengths， existing license plate recognition models generally have insufficient general adaptability and cannot balance recognition accuracy and real-time speed. Severe shortage of samples for two-row structured， complex-background and variable-length character license plates further hinders the development of general-purpose license plate recognition research. To solve these problems， this paper proposes a wide-range license plate recognition network PL-TransLPRNet based on a lightweight parallel encoder-decoder architecture， and a multi-type license plate data generation scheme with a redundancy recycling strategy.MethodTransLPRNet is composed of a lightweight visual encoder， a modal adapter and a parallel decoder based on learnable query tokens. In the lightweight visual encoder， two-stage cascaded high-channel downsampling-free inverted residual modules combined with full residual connections are adopted. Features from initial downsampling convolution are stacked layer-by-layer with two-stage enhanced features to achieve efficient extraction and aggregation of three-level progressive multi-scale features of license plate characters. On this basis， a third-stage downsampling inverted residual module is introduced， and a residual branch based on max-pooling is embedded to make up for the loss of key stroke information caused by downsampling. The extracted features are then input into a refined lightweight MobileViT module. Parallel grouped Transformer layers realize parallel interactive modeling of local and global features， which are aggregated by an adapter and then sent to a parallel Transformer decoder. This decoder abandons traditional autoregressive sequence dependency and adopts a one-step parallel decoding strategy with two cascaded decoding layers. In the first layer， learnable pattern query vectors and position query vectors are used to construct query tokens； through self-attention and cross-attention， coarse matching results are obtained. In the second layer， the coarse results are refined via self-attention for error correction and then cross-attention for fine matching， followed by a feed-forward network and a linear classifier to output class probability distributions for each character position in a batch decoding manner. Thus， the decoder adaptively identifies various license plates with variable layouts and character lengths. Aiming at the scarcity of two-row license plate data， a redundancy recycling strategy is proposed. Easily recognizable license plates from single-row test datasets are recycled as redundant samples. Virtual target license plates are programmatically generated and geometrically fitted by using their real backgrounds and geometric parameters of license plate regions. Without expanding dataset scale or generating extra collection costs， this strategy provides reliable data support for the training and verification of general-purpose license plate recognition models， and further constructs a hybrid dataset with both real and virtual samples covering diverse plate types.ResultsBased on the widely used CCPD dataset， two hybrid real-virtual datasets are constructed with real backgrounds and geometric parameters of license plate regions of redundant samples from its test set. One dataset contains 7-digit and 8-digit single- and two-row Chinese license plates， and the other covers 6-digit and 7-digit New York State license plates as well as 7-digit Chinese license plates. Models trained on the three datasets are comprehensively compared with 10 open-source and 8 closed-source license plate recognition algorithms. To compare fairly with closed-source algorithms， four license plate region extraction methods are applied to reduce interference brought by different license plate localization algorithms. Cross-dataset evaluation is carried out with the CLPD and PKUdata datasets to test the cross-dataset recognition performance of the proposed model. Experimental results show that the proposed model significantly outperforms the comparison algorithms in terms of generalization performance， recognition accuracy and inference efficiency. It achieves an average license plate recognition accuracy of 99.05% on the CCPD dataset. The proposed network has only 1.44 M parameters and yields an inference frame rate of 208.6 FPS， demonstrating strong cross-dataset generalization ability.ConclusionThe data augmentation strategy based on redundancy recycling proposed in this paper effectively relieves the shortage of multi-type license plate experimental data. The proposed model solves the problem of unbalanced generalization， accuracy and real-time performance in license plate recognition from the architectural perspective， and has great practical value and deployment potential.关键词：general license plate recognition;open environment;Transformer;parallel encoder-decoder architecture;double-row license plate synthesis65|33|0更新时间：2026-09-15 -
摘要：Semi-supervised learning （SSL） has rapidly evolved into a cornerstone paradigm for medical image segmentation， offering a pragmatic resolution to the chronic scarcity of high-quality annotated data in clinical environments. Unlike fully supervised approaches that demand exhaustive pixel-level labeling， SSL frameworks leverage vast repositories of unlabeled images alongside limited annotated counterparts to achieve competitive， and often near-fully-supervised， segmentation performance. This survey presents a systematic and multidimensional review of semi-supervised medical image segmentation， covering methodological taxonomies， quantitative benchmarking， persistent challenges， and transformative research frontiers.We begin by establishing a hierarchical taxonomy that categorizes existing SSL segmentation methods into three principal streams： pseudo-labeling， consistency regularization， and co-training. For each category， we dissect representative models published at leading conferences and journals—including CVPR， ECCV， ICCV， MICCAI， TMI， MIA， and TPAMI—with an emphasis on their theoretical underpinnings， architectural innovations， and optimization strategies. In the pseudo-labeling domain， we analyze advances in high-confidence pseudo-signal generation， boundary-aware refinement， uncertainty-guided filtering， and label correction via generative diffusion models. For consistency-based methods， we explore how recent studies address multi-domain heterogeneity and complex anatomical structures through multi-level feature alignment， frequency-space consistency， task-affinity regularization， and adversarial perturbation training. In co-training， we examine multi-model collaboration mechanisms， including diverse architecture design， disagreement-based uncertainty estimation， dynamic fusion strategies， and cross-modal synergy. Beyond technical exposition， we provide a granular assessment of the strengths， limitations， and suitable application scenarios associated with each methodological branch.A distinguishing contribution of this review lies in its extensive empirical benchmarking. We compile and evaluate 34 state-of-the-art semi-supervised segmentation models published between 2022 and 2025 across three publicly accessible medical imaging datasets： ACDC， Pancreas-NIH， and LA. Quantitative comparisons are conducted using three widely adopted metrics—Dice， IoU， and 95HD—providing a multi-faceted view of model performance in terms of volumetric overlap， boundary fidelity， and robustness to structural variability. This side-by-side comparison enables identification of performance leaders such as SDCL， FedSemiDG， CMMT-Net， and CV-SSL-MIS， and facilitates analysis of architectural commonalities underpinning their success， including discrepancy-aware correction， domain generalization in federated settings， cross-head mutual learning， and dual-contrastive transformer backbones. Through this analysis， we illuminate the evolving performance landscape and reveal persistent gaps in cross-modal generalization and small-object segmentation.We further consolidate and characterize ten commonly adopted datasets in the field—including ACDC， PROMISE12， M&Ms， BUSI， GlaS， MoNuSeg， LA， Pancreas-NIH， NCI-ISBI13， and I2CVB—detailing their imaging modalities， annotation protocols， dataset scales， and resolution characteristics. This resource-oriented summary provides practical guidance for researchers selecting appropriate benchmarks for method validation or domain transfer experiments.Despite considerable progress， semi-supervised medical image segmentation remains confronted by several systemic challenges. We critically examine four core dilemmas. First， pseudo-label quality control remains precarious； noisy， overconfident， or anatomically implausible pseudo-labels can propagate errors and destabilize training. We review mitigation strategies including uncertainty-based filtering， temporal ensembling， cooperative rectification networks， and diffusion-based label refinement. Second， class imbalance—particularly pronounced for small lesions， organs with low volumetric occupancy， or rare pathological patterns—is exacerbated in SSL settings due to biased pseudo-labeling toward majority classes. We discuss class-sensitive loss functions， re-sampling strategies， generative augmentation， and balanced confidence calibration as emerging countermeasures. Third， data privacy and regulatory constraints hinder the aggregation of distributed medical data， challenging traditional centralized SSL paradigms. We examine privacy-preserving alternatives， with particular attention to federated semi-supervised learning， differential privacy， and their intersection with domain generalization and communication-efficient optimization. Fourth， we identify cross-domain and cross-modal generalization as a fundamental bottleneck， especially when training and testing distributions diverge due to differences in scanner protocols， patient populations， or imaging modalities. We summarize recent efforts in domain adaptation， modality-collaborative learning， and invariant feature disentanglement.Looking forward， we articulate five forward-looking research directions poised to redefine the landscape of semi-supervised medical image segmentation. 1） Medical knowledge-driven segmentation seeks to transcend purely data-driven representations by embedding explicit anatomical priors， geometric constraints， or radiological semantics into SSL pipelines. We review emerging paradigms that encode shape priors via adversarial learning， distill textbook knowledge via vision-language pretraining， and employ large language models to generate dynamic prompts or pseudo-labels. However， we also caution against over-rigid knowledge injection that may conflict with image evidence， advocating for adaptive fusion mechanisms and retrieval-augmented frameworks. 2） Visual foundation models， particularly the Segment Anything Model （SAM） and its medical adaptations （e.g.， MedSAM， SAM2）， present unprecedented opportunities for zero-shot and few-shot segmentation. We analyze how SAM can be integrated into SSL frameworks through parameter-efficient fine-tuning， automated prompt generation via multimodal LLMs， video-inspired volumetric segmentation， and teacher-student co-training. Critical open issues—domain gaps between natural and medical images， computational overhead， and knowledge conflict—are examined with proposed mitigation pathways. 3） Partial annotation in 3D volumetric segmentation is gaining urgency as fully labeling 3D medical scans remains prohibitively labor-intensive. We highlight recent advances leveraging sparsely labeled slices， orthogonal plane annotations， or scribble-based supervision， combined with multi-level consistency， signed distance function constraints， and pseudo-3D to true-3D progressive learning. Adaptive annotation selection via active learning and uncertainty sampling is identified as a promising synergy. 4） Federated semi-supervised learning offers a privacy-compliant framework for multi-institutional collaboration without centralized data sharing. We analyze recent contributions in dynamic client weighting， semi-centralized training paradigms， and frequency-space domain generalization. We further identify underexplored areas including incentive mechanisms via blockchain， neural architecture search for heterogeneous client capabilities， and personalized federated tuning. 5） Lightweight and efficient models are imperative for clinical deployment on edge devices， intraoperative systems， and low-resource settings. We survey recent breakthroughs in state-space models （e.g.， Mamba-based architectures）， multi-scale dynamic feature fusion， and hardware-aware neural architecture search that substantially reduce parameter counts while preserving segmentation fidelity. Challenges in maintaining cross-modal robustness and achieving real-time inference are critically assessed.In summary， this review provides a comprehensive， empirically grounded， and forward-looking synthesis of semi-supervised medical image segmentation. By bridging methodological innovation， rigorous benchmarking， dataset characterization， challenge dissection， and future roadmap， we aim to equip researchers and practitioners with both a holistic understanding of the current landscape and actionable insights for advancing the field toward greater accuracy， generalizability， privacy compliance， and clinical translatability.关键词：semi-supervised learning;medical image segmentation;Pseudo-labeling;consistency regularization;Co-training52|39|0更新时间：2026-09-15 -
摘要：Military vehicle detection in high-resolution optical remote sensing imagery is a critical link in battlefield situational awareness and intelligence acquisition， and it imposes stringent requirements on both timeliness and accuracy. In remote sensing images， military vehicle targets generally appear at small scale， in dense distributions and with arbitrary orientations， while realistic reconnaissance scenarios are frequently accompanied by complex background clutter and occlusion， artificial camouflage interference， and adverse weather such as haze and fog. These factors are strongly coupled： the physical characteristics of the target itself， its interaction with the surrounding environment， and the imaging conditions act jointly to degrade the observed imagery， so that the distinguishability of a target from its background and the separability among neighboring targets are both markedly reduced. Under foggy conditions in particular， atmospheric scattering attenuates the radiance reflected from the scene and superimposes an additive veil of atmospheric light， and this attenuation grows rapidly with scene depth. As a result， image contrast and edge energy decay severely， directly eroding the contour cues on which the detection of small vehicles most heavily relies. Optical remote sensing detection of military vehicles therefore remains an extremely challenging task.Existing studies have addressed this problem mainly along two lines. For generic detection under complex backgrounds， attention mechanisms and context modeling have been widely adopted to suppress background interference and to compensate for the feature loss caused by occlusion， while saliency modeling and generative adversarial network based approaches have shown potential for enhancing camouflaged and low-resolution targets. For detection under foggy weather， cascaded schemes that first dehaze and then detect can improve accuracy， but they suffer from high computational complexity and error accumulation， because the dehazing stage is optimized for perceptual quality rather than for downstream detection. Parallel dual-branch frameworks that jointly optimize restoration and detection achieve end-to-end learning， yet still encounter bottlenecks such as strong dependence on paired hazy–clear training data， complicated network architectures， and the difficulty of balancing the loss weights between the dehazing and detection subtasks. Meanwhile， generic remote sensing detectors are largely built upon natural-image baselines， and their adaptation to the distinctive properties of remote sensing imagery remains limited； the attention modules they employ often incur considerable computational overhead that is hard to reconcile with the real-time demands of military reconnaissance. More fundamentally， the intrinsic prior knowledge carried by vehicle targets in remote sensing images—in particular their rigid geometry and stable aspect-ratio distribution—has not yet been effectively exploited. Overall， current methods still exhibit evident deficiencies in adapting to remote sensing characteristics， in balancing efficiency against accuracy， and in generalizing to real-world scenarios.Motivated by these observations， and taking YOLOv11n（you only look once）， which offers a favorable trade-off between detection accuracy and inference speed， as the baseline， this paper proposes a fog-density-guided detection algorithm for military vehicles in real-scenario remote sensing imagery， termed MVF-YOLO （YOLO for military vehicle detection under foggy conditions）. The method improves the baseline at the input， the neck and the detection head， and is supported by a purpose-built dataset. At the network input， a fog-density-driven edge restoration module （FDER） is introduced. It fuses the dark channel prior with multi-scale gradient features to estimate a pixel-wise fog density map that characterizes the spatially non-uniform degradation of the scene， and then uses this map to adaptively re-weight and restore the edge responses attenuated by scattering， so that heavily fogged regions receive stronger edge compensation while clear regions remain essentially unchanged and are not over-sharpened. Because FDER carries out edge recovery at the feature level rather than performing image-level dehazing， it requires neither paired hazy–clear supervision nor an additional restoration branch， and it improves feature quality under degraded imaging conditions at very low cost. In the neck， a spatial attention-based frequency selection module （SFS） is designed. It first employs spatial attention to guide the network toward the key regions likely to contain targets， and then performs selective processing in the frequency domain， suppressing low-frequency components that mainly carry smooth background， haze veil and residual noise， and enhancing high-frequency components that encode contours， textures and structural discontinuities—the cues that are most discriminative for rigid vehicle targets. This joint spatial–frequency selection strengthens the recognition of occluded and camouflaged targets whose responses would otherwise be submerged in background energy. In the detection head， a classification loss based on the aspect-ratio prior of the target， denoted ar_BCELoss， is proposed. Exploiting the rigid geometric nature of military vehicles， whose aspect ratios are concentrated within a stable interval， the loss modulates the standard binary cross-entropy according to the consistency between the geometry of a candidate and the prior distribution， thereby emphasizing samples that conform to the vehicle shape prior and down-weighting background responses with implausible geometry； the classification boundary is refined and classification accuracy improved with almost no additional computational overhead. Finally， MVRSD， the only publicly available remote sensing dataset for military vehicle detection in China， is improved and extended： style transfer techniques are used to synthesize foggy samples with realistic and diverse fog densities， yielding the MVRSD-V2.0 dataset， which better represents real-world reconnaissance conditions.Experiments on MVRSD-V2.0 show that the proposed method attains an average precision of 87.6% at an IoU threshold of 0.5， an improvement of 4.4 percentage points over the baseline. Ablation studies confirm that each of the three components contributes positively and that their combination performs best： FDER mainly recovers the accuracy lost on heavily degraded foggy samples， SFS mainly benefits occluded and camouflaged targets in cluttered backgrounds， and ar_BCELoss mainly reduces false alarms caused by background structures with vehicle-like appearance. Since the added modules introduce only marginal growth in parameters and computation， these gains are obtained while the model retains the lightweight nature and inference speed of the YOLOv11n baseline. Comparisons with mainstream detectors and with cascaded dehazing-then-detection pipelines further indicate that MVF-YOLO achieves a better accuracy–efficiency trade-off without relying on an explicit dehazing stage or on paired hazy–clear training data. The experimental results demonstrate that， by jointly performing fog-density-driven edge restoration， spatially guided frequency selection and shape-prior-constrained classification， the proposed method possesses strong robustness and reliable military vehicle detection capability under the coupled degradations of fog， occlusion and camouflage encountered in real scenarios， providing a practical technical reference for remote sensing reconnaissance applications with real-time requirements.关键词：remote sensing imagery;military vehicles;object detection;real-world scenarios;fog density31|35|0更新时间：2026-09-15 -
摘要：ObjectiveSemi-supervised semantic segmentation aims to learn robust pixel-level representations by leveraging a small amount of labeled data and a large amount of unlabeled data， thereby reducing the cost of acquiring high-quality pixel-level annotations. Although existing methods have achieved significant progress under the framework of consistency regularization， most of them still rely on traditional ResNet-based encoders， which limits their feature representation capability. Moreover， existing methods usually impose uniform consistency constraints on the entire image， ignoring the spatial and frequency-domain differences between semantic boundary regions and target interior regions. This may lead to blurred boundaries and inconsistent predictions within object regions. To address these issues， this paper proposes SF-Match， a semi-supervised semantic segmentation framework driven by spatial-frequency consistency.MethodFirst， the traditional ResNet encoder is replaced with DINOv2， a Vision Transformer-based self-supervised pre-trained model. A DPT decoder is further introduced to reconstruct spatial feature maps from sequential representations， forming a dense prediction architecture that is more suitable for semi-supervised semantic segmentation. This architecture enhances the model’s ability to jointly capture global semantic information and local structural details. Based on this architecture， a spatial-frequency consistency mechanism is designed. Specifically， strongly augmented views of unlabeled images are explicitly decomposed into high-frequency and low-frequency views. Meanwhile， morphological gradients are computed from the pseudo-labels generated by the teacher model to dynamically construct boundary-region masks and interior-region masks. High-frequency consistency constraints are applied to boundary regions to strengthen the modeling of object contours and fine-grained details， whereas low-frequency consistency constraints are applied to interior regions to improve structural coherence and class-level consistency. In this way， frequency-domain information and spatial regions are decoupled and aligned in a region-adaptive manner.ResultExperimental results on two benchmark datasets， PASCAL VOC 2012 and Cityscapes， demonstrate the effectiveness of the proposed method. On the PASCAL VOC 2012 dataset， compared with the second-best method， semiVL， SF-Match improves the mean Intersection over Union （mIoU） by 1.9 to 3.6 percentage points under different labeled-data settings. Compared with the strong baseline UniMatch， SF-Match achieves consistent mIoU improvements across all partition settings， with a maximum gain of 11.2 percentage points. On the Cityscapes dataset， compared with the second-best method MPMC， SF-Match improves the mIoU by 5.1 and 3.7 percentage points under the 1/16 and 1/4 labeled-data settings， respectively. In particular， under the 1/16 labeled-data setting， SF-Match achieves an mIoU of 83.6%. Ablation experiments and qualitative analyses further demonstrate that the proposed method effectively improves complex boundary delineation and enhances prediction consistency within object regions.ConclusionThe proposed SF-Match framework constructs a DINOv2–DPT dense prediction network and integrates a spatial-frequency consistency alignment strategy. By jointly enhancing the modeling of global semantic information and fine-grained structural details， SF-Match alleviates the limitations of insufficient feature representation and coarse consistency constraints in existing semi-supervised semantic segmentation methods. Experimental results show that the proposed method achieves superior segmentation performance on multiple benchmark datasets.关键词：semi-supervised learning;semantic segmentation;spatial-frequency consistency;boundary awareness;vision Transformer32|21|0更新时间：2026-09-15 -
摘要：Medical image recognition is a crucial research direction in medical artificial intelligence. It is widely applied in clinical workflows such as computed tomography （CT）， magnetic resonance imaging （MRI）， ultrasound， dermoscopy， X-ray photography， and pathological sections. These workflows involve lesion detection， disease classification， organ segmentation， and auxiliary diagnosis tasks. However， several factors degrade model recognition performance on small lesions， rare diseases， specific patient subgroups， or specific medical centers. These factors include small lesion area proportions， imbalanced disease category distributions （including long-tailed category distributions）， cross-institutional imaging discrepancies， and inconsistent annotation standards. To address these issues， this paper systematically summarizes debiasing methods in medical image recognition. These methods are categorized into four primary classes： pixel-level debiasing methods， sample-level debiasing methods， feature-level debiasing methods， and algorithm-level debiasing methods. Firstly， pixel-level debiasing methods enhance the training contribution and loss weight of small lesions and low-contrast regions in dense prediction tasks， including segmentation and detection. Consequently， small lesion regions receive sufficient attention relative to background regions during network optimization. Pixel-level debiasing methods are categorized into pixel re-weighting debiasing methods， boundary re-weighting debiasing methods， and region re-weighting debiasing methods. Specifically， pixel re-weighting debiasing methods assign different training contributions to individual pixels during pixel-wise loss computation. This strategy weakens the dominant role of massive background pixels in the total loss. Boundary re-weighting debiasing methods introduce ground-truth boundaries， distance fields， or boundary discrepancy regions into the loss function. Thus， models during training not only focus on whether pixel categories are correct， but also pay close attention to spatial offsets between predicted contours and ground-truth contours. Region re-weighting debiasing methods adjust loss contributions based on the overall relationship between predicted regions and ground-truth regions. These adjustments utilize region overlap， category volume， false-detection and missed-detection costs， or region size to alleviate lesion area bias in segmentation tasks. Secondly， sample-level debiasing methods enhance the training contribution of minority-class samples under data imbalance conditions. Sample-level debiasing methods are categorized into sample rebalancing debiasing methods and sample expansion and generation debiasing methods. Specifically， sample rebalancing debiasing methods adjust sample extraction methods， screening methods， or training stage schedules without altering original image content. This improves the effective contribution of minority-class samples during model training. Sample expansion and generation debiasing methods synthesize realistic medical images that preserve semantic information for minority-class samples. This alleviates sample scarcity without violating biological logic or radiological rationality. Thirdly， feature-level debiasing methods resolve uneven cross-domain feature distributions. These distributions are caused by non-disease factors， including hospital sources， equipment types， and demographic group attributes. Feature-level debiasing methods are categorized into feature alignment debiasing methods， feature decoupling debiasing methods， and knowledge distillation debiasing methods. Specifically， feature alignment debiasing methods constrain distribution discrepancies across different domains in high-level feature space. This enables similar diseases across different hospitals， equipment， or datasets to have closer representations. Feature decoupling debiasing methods separately model disease-relevant information from domain styles， color backgrounds， or sensitive attributes during feature learning. Consequently， models rely more on stable disease features. Knowledge distillation debiasing methods utilize teacher models to supervise student models. This design allows student models to absorb more stable category relationships， structural relationships， or fairness information while learning primary tasks. Fourthly， algorithm-level debiasing methods reduce model reliance on non-object attributes during algorithm training. Algorithm-level debiasing methods are summarized as causal constraint debiasing methods. Causal constraint debiasing methods rectify and constrain feature selection from a causal perspective. This guides models to learn true pathological causal relationships and reduces reliance on biased attributes during prediction. Furthermore， this paper analyzes benchmark evaluation experiments across these four categories of debiasing methods. It systematically summarizes commonly used benchmark datasets and evaluation metrics. Finally， this paper summarizes four critical challenges currently confronting fairness research in medical image recognition： accurately identifying bias sources such as hospital origins， equipment types， and demographic group attributes remains difficult； achieving an optimal balance between debiasing and preserving clinically relevant diagnostic information is challenging； the authenticity， biological validity， and privacy compliance of generated medical images still require improvement； and cross-center fairness evaluation systems remain incomplete. This paper provides a valuable reference for the research and application of debiasing methods in medical image recognition.关键词：fairness issue;pixel-level bias;sample-level bias;feature-level bias;algorithm-level bias26|27|0更新时间：2026-09-15 -
摘要：ObjectiveCompositional zero-shot learning （CZSL） aims to recognize unseen state-object compositions by transferring knowledge from seen states， objects and their observed combinations. Unlike conventional closed-set recognition， CZSL requires a model not only to identify primitive concepts， but also to recombine them into novel semantic compositions without direct visual supervision. Recent methods based on vision-language pre-trained models have improved CZSL by using large-scale image-text knowledge， prompt learning and cross-modal alignment. However， two issues remain insufficiently explored. First， visual representations may still be biased toward the color， texture and background statistics of training images， which limits their robustness to unseen compositions. Second， state semantics are often determined by local regions， such as the texture of a wet dog， the surface of a rusty car or the cut surface of a sliced tomato. Global feature interaction may introduce irrelevant background information and weaken local discriminative cues. To address these problems， this paper proposes a CZSL method based on style debiasing and local semantic focusing.MethodThe proposed method is built upon a three-branch vision-language framework， including a composition branch， an attribute branch and an object branch. To improve visual robustness， a two-dimensional style debiasing module is inserted into the visual front end. Specifically， after the convolutional patch embedding layer of the CLIP visual encoder， the channel-wise mean and standard deviation of intermediate feature maps are computed. The feature maps are normalized and then reconstructed with mixed style statistics sampled from different images in the same batch. This operation perturbs feature-level style distributions while preserving the main semantic content， thereby reducing the dependence of the model on training-specific visual styles. To enhance local semantic discrimination， a text-anchored local semantic focusing module is further designed. Text prototypes are generated for the composition， attribute and object branches， and their aggregated representations are used as branch-level semantic anchors. The semantic anchor is matched with visual patch tokens through normalized similarity calculation. A Top-K sparse selection strategy is then adopted to retain the most relevant patch tokens， and the selected tokens are re-weighted according to their semantic relevance. The resulting weighted visual token sequence is used for cross-modal relationship modeling. Finally， the three branches jointly predict composition， attribute and object matching scores， and the final state-object prediction is obtained by fusing these scores.ResultExperiments are conducted on three benchmark datasets， including MIT-States， UT-Zappos and C-GQA， under both closed-world and open-world settings. Four metrics are used for evaluation： Seen accuracy （S）， Unseen accuracy （U）， Harmonic Mean （H） and Area Under Curve （AUC）. In the closed-world setting， the proposed method achieves 50.9% S， 53.7% U， 39.8% H and 22.9% AUC on MIT-States； 67.6% S， 74.4% U， 55.3% H and 42.8% AUC on UT-Zappos； and 41.7% S， 36.4% U， 30.2% H and 13.1% AUC on C-GQA. In the open-world setting， it achieves 49.8% S， 20.3% U， 20.9% H and 7.8% AUC on MIT-States； 67.4% S， 62.1% U， 48.4% H and 33.8% AUC on UT-Zappos； and 41.6% S， 8.3% U， 11.7% H and 3.24% AUC on C-GQA. Compared with the baseline three-branch model， the proposed method consistently improves performance on all three datasets， especially on H and AUC， indicating a better balance between seen and unseen compositions. Ablation experiments on MIT-States show that adding only the style debiasing module improves the closed-world AUC from 22.1% to 22.4%， adding only the local semantic focusing module improves it to 22.6%， and using both modules further improves it to 22.9%. These results verify the complementary effects of the two modules.ConclusionThis paper presents a compositional zero-shot learning method that combines style debiasing and local semantic focusing. The style debiasing module alleviates the influence of training-specific visual statistics， while the text-anchored local semantic focusing module enhances the selection of discriminative patch-level evidence. By integrating these modules into a three-branch cross-modal framework， the proposed method improves both visual robustness and local semantic discrimination. Experimental results， ablation studies and qualitative analysis demonstrate its effectiveness in recognizing unseen state-object compositions. Future work will explore more fine-grained region-level semantic grounding and stronger external knowledge constraints for difficult compositions with subtle states， weak local cues or complex backgrounds.关键词：Compositional Zero-Shot Learning;vision-language model;style debiasing;local semantic focusing;cross-modal matching7|7|0更新时间：2026-09-15 -
摘要：ObjectiveContinuous action segmentation and prediction are important for human-computer interaction， intelligent perception， and surgical activity understanding. Continuous action videos usually contain rapid local motion changes， long-range temporal dependencies， ambiguous transitions， and large variations in action duration. Although existing methods based on three-dimensional convolutional networks， recurrent networks， temporal convolutional networks， or Transformer structures have improved spatiotemporal representation learning， they still have difficulty in jointly modeling short-term dynamic details and long-term semantic context. Weak boundary cues in transition regions may also cause temporal jitter， over-segmentation， and inaccurate boundary localization. For action prediction， future categories must be inferred from partially observed sequences， which further requires stable temporal modeling and effective representation of motion trends. To address these issues， a video action segmentation and prediction method is proposed based on cross-scale temporal interaction and boundary probability refinement， aiming to improve discriminative representation， temporal consistency， boundary localization， and future prediction stability.MethodThe proposed method establishes a unified framework consisting of local spatiotemporal encoding， cross-scale temporal interaction， and boundary probability refinement. For action segmentation， RGB video sequences are encoded by an I3D network to obtain frame-level spatiotemporal features. For action prediction， RGB and optical flow sequences are used as dual-modal inputs， and an X3D network is used to extract features from both modalities. A bidirectional cross-attention mechanism enhances multimodal interaction， in which RGB features aggregate motion information from optical flow features and optical flow features obtain complementary appearance information from RGB features. A gated fusion strategy integrates the enhanced RGB and flow representations. Based on the extracted feature sequence， the cross-scale temporal interaction module adopts an encoder-decoder structure to model temporal dependencies at different hierarchical levels. The encoder aggregates global contextual information， whereas the decoder fuses high-level context with lower-level temporal details， thereby preserving both long-range dependencies and fine-grained local dynamics. The boundary probability refinement module explicitly models action transition boundaries through refinement stages based on dilated convolution and residual connections. The refined features are mapped to frame-level boundary probabilities. A weighted binary cross-entropy loss is used for boundary supervision， and a boundary-gated temporal consistency constraint suppresses unnecessary prediction fluctuations within action segments while allowing category changes near true boundaries. The whole framework is optimized by classification loss， boundary prediction loss， and boundary-gated temporal consistency loss.ResultExperiments were conducted on four public continuous action segmentation datasets， including 50Salads， GTEA， Breakfast， and JIGSAWS. Frame-wise accuracy， edit score， and F1 scores under different intersection-over-union thresholds were used as evaluation metrics. On 50Salads， the proposed method achieved 93.6， 92.2， and 88.8 in F1@10， F1@25， and F1@50， respectively， with an edit score of 89.0 and an accuracy of 92.3. On GTEA， the method obtained the best results in F1@10， F1@25， edit score， and accuracy， indicating its effectiveness for fine-grained first-person action segmentation. On Breakfast， the method achieved 86.4 in F1@10， 82.5 in F1@25， 84.7 in edit score， and 83.2 in accuracy， demonstrating robustness under complex long-sequence conditions. On JIGSAWS， it achieved 97.1， 95.7， and 90.6 in F1@10， F1@25， and F1@50， respectively， with an edit score of 95.2 and an accuracy of 90.5， suggesting that it can effectively model fine-grained surgical action sequences. For action prediction， the method achieved an accuracy of 90.3 on JIGSAWS， outperforming representative prediction methods. Ablation experiments confirmed the contribution of each key component. Removing the cross-scale temporal interaction module caused performance degradation on all datasets， especially in F1@50 and edit score， indicating that cross-scale temporal modeling improves boundary recovery and sequence-level consistency. Removing the boundary probability refinement module also reduced performance， particularly in high-threshold F1 and edit score， verifying that explicit boundary modeling can reduce over-segmentation and improve transition localization. RGB and optical flow also showed complementary effects in action prediction， and their fusion achieved higher accuracy than either modality alone.ConclusionThe proposed method integrates local spatiotemporal encoding， cross-scale temporal interaction， and boundary probability refinement into a unified framework for continuous video action segmentation and prediction. By combining short-term dynamic information with long-range temporal context， the model enhances the representation of action evolution across different temporal scales. By explicitly learning boundary probabilities and introducing boundary-gated temporal consistency constraints， it improves frame-level classification stability and action boundary localization. For action prediction， the fusion of RGB appearance information and optical flow motion information strengthens the perception of future action trends. Experimental results on multiple public datasets demonstrate that the proposed method achieves competitive or superior performance compared with existing methods， especially in temporal consistency， boundary localization， and future action prediction. The method provides an effective solution for continuous action understanding in complex video scenarios.关键词：continuous action segmentation and prediction;cross-scale temporal interaction;boundary probability refinement;spatiotemporal feature encoding;multimodal fusion11|7|0更新时间：2026-09-15 -
摘要：ObjectiveHyperspectral images （HSIs） capture dozens or even hundreds of narrow spectral bands， providing a wealth of spectral details that make it possible to differentiate land-cover classes even when their spectra differ only slightly. With the rapid development of the low-altitude economy， HSIs have been widely used in remote sensing applications such as land-cover mapping， agricultural monitoring， environmental analysis， and urban observation. However， the high dimensionality of hyperspectral data also introduces severe redundancy among adjacent bands， which substantially increases the burden of storage， transmission， and computation in subsequent image analysis. Band selection has become an important dimensionality reduction approach to alleviate these problems. It extracts a compact subset of informative bands directly from the original spectral space while preserving the physical meaning of the selected wavelengths. Existing band selection methods can generally be divided into unsupervised， supervised， and semi-supervised methods. Among these， semi-supervised band selection is particularly attractive because it attempts to take advantage of both labeled and unlabeled samples. However， most existing semi-supervised methods are built for a single dataset， meaning they mainly depend on the small number of labeled samples available within one image. In real-world scenarios， hyperspectral labeling is expensive and difficult to obtain， restricting the discriminative capability of the selected band subsets. In fact， hyperspectral datasets acquired by the same sensor typically exhibit similar physical properties， such as reflectance or radiance characteristics， even though they may correspond to different scenes and land-cover distributions. This indicates that multiple related datasets may contain latent common spectral knowledge that can be jointly explored. Motivated by this， the paper seeks to enhance band selection performance under limited supervision by moving away from optimization on a single dataset and instead adopting collaborative analysis across multiple datasets.MethodTo achieve this goal， an evolutionary multitasking optimization （EMTO）-based multi-dataset collaborative band selection method， termed EMDCBS， is proposed. The core idea is to jointly exploit dataset-specific spectral information and shared discriminative knowledge within a unified multitasking optimization framework. Specifically， for each hyperspectral dataset， EMDCBS establishes a private unsupervised band selection task to search for band subsets with rich spectral information. In parallel， a shared supervised band selection task is constructed by integrating the labeled samples from multiple datasets so as to alleviate the limitation caused by insufficient annotations in any single dataset and to enhance the discriminative capability of the resulting band subset. To optimize these tasks collaboratively， EMDCBS adopts an evolutionary multitasking optimization mechanism in which candidate band subsets are encoded as binary individuals and iteratively evolved through crossover and mutation. During this process， both intra-task interaction and inter-task interaction are introduced. The former enhances the search within the same task， while the latter enables knowledge transfer across different tasks， thereby facilitating the reuse of potentially effective band combinations and improving global search ability. For the private unsupervised tasks， information entropy is used as the objective function to preserve information-rich bands. For the shared supervised task， a separability criterion based on between-class scatter and within-class scatter is adopted to characterize the discriminative quality of the selected bands more directly under limited labeled samples. After task optimization， EMDCBS obtains multiple diverse optimal band subsets from different tasks. Since these subsets emphasize different properties， a multiple band subset fusion strategy is further introduced. By jointly considering information content， discriminative ability， and redundancy， the proposed strategy generates a final band subset with better comprehensive performance for each dataset.ResultThis paper evaluates the effectiveness of the EMDCBS method using four hyperspectral datasets. The overall accuracy （OA） values of the proposed EMDCBS method on the Pavia University， Pavia Center， LongKou， and HongHu datasets are 90.05%， 95.89%， 93.39%， and 79.21%， respectively. These results demonstrate that， compared to other band selection methods， the EMDCBS method can generate a better band subset with limited labeled samples， and achieves better classification performance. Notably， the improvements observed on the LongKou and HongHu datasets indicate that the advantages of collaborative learning across multiple datasets are more pronounced when labeled data are scarce and scene complexity is high. Furthermore， in the parametric analysis， the classification accuracy of all band selection methods increases with the number of labeled samples. However， when the number of labeled samples per class exceeds 50， the improvement in accuracy for all methods tends to plateau. In addition， the impact of the size of the selected band subset on classification performance is also analyzed in the experiments. Experimental results show that increasing the number of selected bands improves classification accuracy， but the performance improvement becomes minimal when the subset size exceeds 20 bands. This indicates that selecting additional bands may introduce redundancy and increase storage and computational burden， but has limited impact on improving classification accuracy. Ablation experiments， by comparing the classification accuracy of single-dataset band selection with multi-dataset collaborative band selection， confirm the effectiveness of multi-dataset collaborative analysis. Furthermore， comparing convergence curves with and without task interaction mechanisms shows that these mechanisms not only accelerate convergence but also produce better band subsets. This indicates that knowledge sharing between tasks positively impacts the evolutionary optimization process. In addition， ablation experiments verify that by balancing information richness and class discriminative ability and reducing redundancy， the multi-band subset fusion strategy can obtain better band subsets. In conclusion， these results confirm that each major component of EMDCBS makes a significant contribution to performance improvement.ConclusionThis study proposes a collaborative semi-supervised hyperspectral band selection approach based on evolutionary multi-task optimization， namely EMDCBS. Unlike traditional approaches that optimize band subsets independently within single datasets， EMDCBS concurrently formulates private unsupervised tasks alongside shared supervised tasks across multiple datasets. EMDCBS employs a collaborative multi-task evolutionary search combined with a multiple band subset fusion strategy. This improves both optimization efficiency and the quality of selected band subsets. As a result， it enables more informative and discriminative band selection even when label supervision is limited. Experimental results conducted on four hyperspectral datasets demonstrate that the proposed method outperforms compared techniques. Nonetheless， several areas for further enhancement remain. First， acknowledging that shared spectral knowledge among tasks is inherently dynamic rather than static， there is a need to develop an adaptive knowledge transfer mechanism that accounts for task relevance， population diversity， and optimization progress to enable more precise and effective inter-task collaboration. Furthermore， with the increase in the number of tasks and the scale of data， the time and computational resources required for evolutionary strategies will increase significantly， making it crucial to explore more efficient search strategies.关键词：hyperspectral images;band selection;semi-supervised learning;multi-dataset collaborative learning;evolutionary multitasking optimization;low-altitude economy7|3|0更新时间：2026-09-15 -
摘要：ObjectiveWith the rapid development and extensive application of unmanned aerial vehicles （UAVs） in both civilian and military fields， all-weather， robust multi-object tracking （MOT） technology has become a core technical requirement for low-altitude security and defense fields. Infrared thermal imaging is capable of stable perception under low-light and all-day conditions， making it the preferred sensing method for UAV monitoring. However， infrared UAV targets are characterized by small pixel scales， low signal-to-clutter ratio （SCR）， high flight speed and strong nonlinear non-stationary maneuvers， which pose great challenges to existing MOT algorithms. Existing open-source infrared anti-UAV datasets generally exhibit obvious evaluation bias. They place excessive emphasis on object detection performance in low SCR scenarios and UAV motion patterns are predominantly stable， with mild， regular trajectories. Such datasets fail to reflect the high-speed variable motion， sharp turning， hovering and complex tactical maneuver behaviors of real UAVs and cannot quantitatively evaluate the robustness of tracking algorithms in highly dynamic and complex environments. In addition， the low SCR of existing datasets mixes detection errors with data association errors， making it difficult to locate the real causes of tracking performance degradation. To address the limitations of existing benchmarks， we construct MIATrack （Maneuver Infrared Anti-UAV Tracking Dataset）， a synthetic infrared anti-UAV MOT dataset for highly maneuverable UAVs. It further reveals the real performance bottleneck of tracking algorithms under complex nonlinear maneuvers.MethodWe first build a library of clean infrared backgrounds covering typical scenes， including skies， sea surfaces， deserts and urban areas， as well as a library of targets featuring various types of UAVs. Based on raw infrared images and mask annotations from public infrared anti-UAV datasets， we remove UAV targets using mask expansion and K-nearest-neighbor inpainting to produce a purified background library. Corresponding UAV instances are also extracted to establish a high-quality infrared UAV target library. We then adopt a physics-driven modeling framework to synthesize realistic infrared image sequences. Multiple kinematic models are employed to create complex flight trajectories， including constant-velocity motion， constant-acceleration motion， circular motion， serpentine maneuvers and Z-shaped zigzag motion. Manual trajectory customization is also supported. We also simulate realistic infrared target characteristics， including distance-dependent geometric scaling， atmospheric radiometric attenuation， attitude-induced coordinate rotation and point-spread-function-based optical blurring. By adjusting generation parameters， MIATrack maintains a high SCR to mitigate detection uncertainty， thereby shifting the focus from detection accuracy to tracking robustness. We adopt adjacent-frame IoU and the relative-position-switching frequency to quantitatively evaluate the motion complexity of MIATrack and conduct comprehensive benchmark tests on the YOLO26 detection model series and representative tracking algorithms， including SORT， ByteTrack， OC-SORT， SparseTrack and ImprAsso.ResultStatistical results show that MIATrack achieves a considerably higher average SCR than existing mainstream datasets， including AntiUAV and 4th-AntiUAV， reflecting its intentional design to maintain cleaner target signatures and reduce detection uncertainty. In terms of motion characteristics， MIATrack presents a markedly lower mean adjacent-frame IoU and a notably higher relative position switching frequency， confirming that it contains highly nonlinear， non-stationary trajectories and complex target interactions. Detection experiments show that YOLO26 models of varying scales achieve favorable mAP on MIATrack， demonstrating that the high SCR design effectively suppresses clutter interference. In MOT tests， mainstream trackers maintain competitive MOTA scores thanks to high-quality detection results. However， they experience a sharp drop in core association metrics， including HOTA， IDF1， and AssA. This pronounced performance discrepancy between detection and association clearly indicates that complex maneuvers severely degrade the trajectory-matching and identity-preservation capabilities of existing tracking algorithms.ConclusionMIATrack compensates for the insufficient motion diversity of existing infrared anti-UAV benchmarks and enables comprehensive and rigorous evaluation of multi-target tracking algorithms in complex maneuver scenarios. Benchmark results indicate that the core bottleneck of infrared anti-UAV MOT has shifted from target detection to motion prediction and data association in complex maneuver scenarios. By maintaining a high signal-to-clutter ratio， MIATrack reduces detection uncertainty， enabling a more rigorous evaluation of robustness in motion modeling and data association under complex maneuvers. Accordingly， this work establishes a new baseline for future tracking algorithm development. The dataset and source code are publicly released to support the design， testing and practical deployment of infrared anti-UAV detection and multi-object tracking methods.关键词：Infrared small target detection and tracking;Maneuver target tracking;multi target tracking;Infrared datasets of anti-UAV;Sequence images78|56|0更新时间：2026-09-07 -
摘要：ObjectivePear rust is a common foliar disease that can reduce photosynthetic capacity， accelerate premature leaf fall， and ultimately affect fruit quality and yield. Unmanned aerial vehicle （UAV） imagery provides an efficient means of surveying orchard canopies over large areas， but reliable automatic detection remains difficult under natural acquisition conditions. Rust lesions are often small and locally similar to background patterns， while infected leaves frequently co-occur with shadows， highlights， white veins， occlusion boundaries， and branch textures. These factors may become spuriously correlated with disease labels during detector training and can produce high-confidence false positives. Existing agricultural detectors mainly improve feature extraction， multi-scale fusion， attention， or detection heads， but seldom model the confounding risk of individual prediction queries. This study therefore proposes a causal-inspired deconfounding method for pear rust detection in UAV imagery， aiming to reduce the influence of background-associated spurious cues on detection confidence while preserving the original localization capability of the detector. Method Real-Time Detection Transformer （RT-DETR） is adopted as the baseline detector because its Transformer decoder represents each candidate prediction with an explicit object query that is naturally aligned with the final classification and box-regression outputs. The proposed framework introduces a lightweight causal confounder discriminative filtering module， referred to as CCDF， after the baseline detector. First， query-level target and confounding proxies are constructed using a frozen RT-DETR. For each image， the 256-dimensional hidden query representation output by the last Transformer decoder layer and fed into the final classification prediction head is extracted before classification logits are produced. This level is selected because it has undergone the complete decoder updates and corresponds directly to the final candidate predictions. Earlier decoder-layer representations could be used， whereas encoder or backbone features would require additional candidate-region alignment and pooling. Changing the extraction layer would require reconstructing the proxy sets and retraining the discriminator； cross-layer performance differences are not quantified here. Candidate queries are divided according to their original confidence and maximum intersection-over-union （IoU） with ground-truth boxes. Queries with confidence not lower than 0.25 and IoU not lower than 0.50 are used as target-evidence proxies， whereas predictions with confidence not lower than 0.50 and maximum IoU lower than 0.10 are used as confounding proxies. The interval between the two IoU criteria is excluded to reduce label ambiguity caused by localization errors around small lesions. The thresholds are fixed to balance proxy purity and sample availability； adaptive thresholds remain feasible but require validation. A lightweight 256-128-2 multilayer perceptron （MLP） is then trained on the two proxy sets to learn a nonlinear boundary between target-evidence and confounding patterns and to output a query-level confounding probability. Finally， the estimated probability is converted into a continuous risk weight and used in a monotonic multiplicative re-scoring function. The mechanism suppresses candidates with higher estimated confounding risk more strongly while relatively preserving lower-risk candidates. It changes only confidence values and their ranking， without modifying the predicted category， bounding-box coordinates， or RT-DETR parameters. Three suppression configurations are evaluated to characterize the precision-recall trade-off， with strategy A used as the main configuration for stronger false-positive control. To avoid information leakage， proxy mining and discriminator training use only the training split； the validation split is used for discriminator evaluation and re-scoring configuration selection， while the test set is reserved for final performance evaluation and visualization. Ground-truth annotations from the test set are used only after inference to compute evaluation metrics and are never provided to CCDF during prediction. Results Experiments are conducted on the public GYMNSA pear rust detection dataset using a unified test set containing 140 UAV images and 2，135 annotated rust instances. The baseline RT-DETR is trained for 100 epochs under the same evaluation protocol. On the held-out validation data， the MLP confounding discriminator achieves an area under the receiver operating characteristic curve （AUROC） of 0.813， an area under the precision-recall curve （AUPRC） of 0.971， and an average score-separation value of 0.363， indicating that target-evidence and confounding queries are distinguishable in the selected query representation space and that the discriminator output can serve as a meaningful risk variable for subsequent re-scoring. On the GYMNSA test set， the RT-DETR baseline obtains a Precision of 0.755， a Recall of 0.664， a mean average precision at IoU=0.50 （mAP@50） of 0.722， and 460 false-positive detections. With the proposed CCDF and the main suppression strategy， Precision increases to 0.781 and mAP@50 reaches 0.723， while the number of false positives decreases to 348. This corresponds to 112 fewer false positives， or a reduction of 24.35%， whereas Recall decreases to 0.641. This is consistent with the intended role of CCDF： it does not generate boxes or alter regression， but preferentially down-weights predictions resembling the confounding proxy distribution. Ablation experiments with three suppression intensities show a systematic precision-recall trade-off. Stronger suppression produces higher Precision and fewer false positives， whereas weaker suppression retains more difficult low-confidence lesions. Candidate-level score analysis further shows that score reduction increases with estimated confounding probability. For high-confidence confounding candidates， the median score reduction is 0.172； 66.73% of such candidates are moved below the decision threshold of 0.55， compared with 23.54% of the corresponding target candidates. Qualitative analysis also shows that many baseline errors occur around reflections， leaf veins， shadows， and occlusion boundaries， and that CCDF effectively removes a substantial portion of these high-confidence predictions from the final detection set. Conclusions The proposed method introduces lightweight causal-inspired deconfounding into an end-to-end Transformer detector without pixel-level confounder masks， domain labels， detector retraining， or changes to bounding-box regression. By treating confident background false positives as observable query-level proxies for confounding effects and well-matched lesion queries as proxies for target evidence， CCDF module converts otherwise difficult-to-observe environmental interference into a supervised risk estimation problem. The learned confounding probability then supports continuous confidence calibration that preferentially attenuates predictions dominated by spurious background patterns. Results demonstrate that this strategy substantially reduces false alarms while maintaining nearly unchanged mAP@50， thereby improving the reliability of UAV-based pear rust screening in scenarios where false positives are costly. The decrease in Recall also indicates that aggressive deconfounding may suppress a subset of difficult true lesions， so CCDF should be understood as an explicit precision-recall control mechanism rather than a universal performance enhancer. Future work will investigate adaptive proxy-selection thresholds， cross-layer feature representations， and end-to-end integration of deconfounding constraints to improve difficult-lesion recall and robustness across orchards， illumination conditions， flight configurations， and disease stages.关键词：pear rust disease detection;causal inference;deconfounding;backdoor adjustment;UAV imagery41|30|0更新时间：2026-09-07 -
摘要：Lithium-ion batteries have become the core energy storage devices in consumer electronics， electric vehicles， and energy storage systems due to their outstanding advantages， such as high energy density， long cycle life， and high charge–discharge efficiency. However， owing to the complex internal structure of batteries， various defects， including electrode and separator defects， are prone to occur during manufacturing. The presence of such defects may induce short circuits， thermal runaway， and other hazardous reactions， which can lead to battery leakage， fire， or even explosion. Non-destructive imaging defect detection approaches have been extensively employed in full-process quality inspection of lithium-ion battery manufacturing owing to their non-destructive nature， high precision， and high efficiency. This paper presents a comprehensive survey of defect detection methods based on both visible-light and non-visible-light imaging for lithium-ion batteries. This review systematically summarizes the working principles， performance characteristics， and applicable scenarios of various detection techniques， and analyzes the existing technical limitations. Visible-light imaging-based defect detection methods for lithium-ion batteries generally adopt artificial light sources coupled with industrial cameras to capture surface images of batteries， and then employ image processing algorithms to identify various surface defects. This represents the most widely used scheme for the inspection of surface defects on lithium-ion batteries. This paper elaborates on three major categories： image processing-based methods， traditional machine learning-based methods， and deep learning-based methods. The core idea of image processing-based methods is to enhance the contrast between defective regions and the background through image preprocessing and feature extraction， thereby realizing defect identification and segmentation. On this basis， this paper further discusses image fusion， to improve the detection capability for tiny and complex defects. Traditional machine learning-based methods incorporate machine learning algorithms on the basis of image processing and construct defect recognition models by extracting handcrafted features， such as texture， shape， and gray-level features， from images to classify and identify defect types. Deep learning-based methods significantly improve the accuracy and efficiency of defect detection， especially for tiny and complex defects. By automatically extracting deep-level defect features from large-scale image datasets without manually designed feature descriptors， these methods enable end-to-end defect identification and segmentation. This paper highlights the application of data augmentation in few-shot scenarios and the deployment of Transformer-based models. Non-visible-light imaging-based defect detection methods for lithium-ion batteries address the limitations of visible-light imaging by employing technologies such as laser imaging， infrared imaging， ultrasonic imaging， computed tomography （CT）， X-ray imaging， and electromagnetic imaging. These modalities enable accurate detection from the external surface to the internal structure of batteries. These imaging modalities provide multi-dimensional information ranging from microstructures to overall morphologies. They compensate for the deficiencies of visible-light imaging and form a complementary relationship with visible-light imaging methods， thereby establishing a comprehensive defect detection system. Laser imaging enables accurate detection of tiny bulges， scratches， and thickness variations on battery surfaces. With high detection precision and strong anti-interference capability， it is suitable for precision defect inspection of battery electrodes， casings， and other components. Nevertheless， it suffers from relatively high equipment costs and low detection speed. Infrared imaging utilizes the fact that internal defects cause local temperature anomalies. By capturing temperature distribution images via infrared thermal imagers， these devices identify internal defects based on temperature differences. When defects such as internal short circuits or separator damage occur， defective regions exhibit abnormal temperatures due to local heat generation， allowing rapid defect localization. Ultrasonic imaging can accurately detect internal defects， including electrode delamination， separator damage， internal voids， and impurities. It exhibits high sensitivity in detecting defects such as internal gassing and lithium plating. However， its penetration ability for high-density materials is limited， and it is susceptible to the complexity of battery internal structures. CT imaging employs multi-angle X-ray scanning combined with computer reconstruction algorithms. This enables clear visualization of subtle internal defects such as electrode misalignment， separator damage， lithium dendrite growth， and internal impurities. It provides extremely high detection precision and supports multi-scale morphological imaging from microscopic features to full-size batteries. X-ray imaging identifies internal battery defects based on X-ray transmission images. It features fast detection speed and online inspection capability， making it suitable for rapid screening of internal defects in large-scale production. Electromagnetic imaging is based on the principle of electromagnetic induction. It identifies conductive and insulation defects inside batteries， such as internal short circuits and current collector fractures， by detecting variations in electromagnetic signals. It offers the advantages of non-contact operation， radiation-free detection， and high detection speed. With the continuously increasing requirements for lithium-ion battery defect detection， 3D imaging-based methods establish three-dimensional models of batteries via 3D imaging technology. This enables the acquisition of three-dimensional features such as defect size and surface topography， thereby addressing the limitations of two-dimensional images in defect characterization. Although visible-light and non-visible-light imaging-based defect detection methods for lithium-ion batteries have achieved remarkable progress and have been widely applied in industrial production， several urgent challenges remain. First， small-sample defect detection capability is insufficient， which limits the generalization ability of deep learning models and makes it difficult to adapt to diverse and complex defect scenarios. Second， multimodal image fusion technology is still immature， hindering the full utilization of the advantages of various imaging modalities. Third， real-time detection performance needs to be improved， as some high-precision detection methods （such as CT and deep learning algorithms） suffer from low computational efficiency. Fourth， the cost of detection equipment is relatively high， especially for non-visible-light imaging and 3D imaging devices. Fifth， unified standards for defect detection and evaluation are still lacking. Finally， based on the above issues， this paper discusses future research directions and puts forward several feasible suggestions.关键词：lithium-ion battery;defect detection;visible light imaging;non-visible light imaging;three-dimensional imaging;image processing methods;deep learning;performance evaluation index43|34|0更新时间：2026-09-07 -
摘要：ObjectiveMountain roads are vital infrastructure for transportation， logistics， and socioeconomic development in remote， topographically challenging regions. They provide essential connectivity for education， healthcare， resource transport， and disaster relief， yet their construction and maintenance are extremely difficult because of steep slopes， unstable geology， and frequent natural hazards. Despite their importance， automatically extracting mountain roads from high-resolution remote-sensing imagery remains a longstanding challenge. Complex terrain introduces severe occlusion and illumination variation， while the spectral appearance of roads is easily confused with bare rock， dry riverbeds， and agricultural paths. In addition， mountain roads are typically narrow， winding， and discontinuous， which differs substantially from the wide， regular patterns found in urban or rural road networks. Existing public road-extraction datasets， such as SpaceNet and DeepGlobe， were developed primarily for flat or urbanized environments. As a result， they fail to capture the spectral， geometric， and contextual complexity of mountain roads， leading to poor generalization when deep-learning models are transferred to high-relief regions. To bridge this gap， we construct AsiaMountain-Road， the first high-resolution， multi-country benchmark dataset dedicated to mountain-road extraction， and propose a novel multi-resolution adaptive feature dual aggregation network （MR-AFDANet） to evaluate and enhance model adaptability in diverse mountainous environments.MethodThe AsiaMountain-Road dataset was built from 29 high-resolution satellite-image areas selected from representative mountainous regions in 17 Asian countries， covering a total of 1 836 km² and providing detailed surface information for fine-scale mapping. All imagery underwent radiometric calibration， orthorectification， and mosaicking， followed by manual annotation by trained experts. Each region was labeled with fine-grained binary road masks to ensure pixel-level accuracy. To improve usability， auxiliary metadata， including elevation maps and land-cover references， were also provided. For algorithmic evaluation， we benchmarked several mainstream deep-learning models—D-LinkNet， OareNet， C2Net， AFDANet， RCFSNet， and RS-Mamba—to assess their adaptability to mountainous terrain. Building on these findings， we propose MR-AFDANet， a model specifically designed for complex terrain mapping. MR-AFDANet uses a ResNet34 encoder as its backbone， consisting of four downsampling convolutional blocks. Each block is followed by a SEAF （spatial-enhanced adaptive feature fusion） module， which adaptively fuses intra-block and residual features while strengthening spatial information through multi-scale convolutions and attention mechanisms. In the decoder， an SRADecoder progressively fuses multi-resolution features to refine road boundaries， preserve continuity in narrow and winding segments， and suppress background noise caused by shadows and vegetation. This design enables the network to preserve thin road topology even under severe spectral interference.ResultExperimental evaluation on the AsiaMountain-Road benchmark demonstrates the substantial difficulty of road extraction in mountainous regions. Compared with urban or plain datasets， mountain imagery contains more complex topography， stronger illumination variation， and more severe occlusion from vegetation and terrain shadows， all of which significantly degrade segmentation accuracy. Across six state-of-the-art baselines， all models show varying degrees of road omission and fragmentation， particularly on narrow， winding， and shaded mountain roads. The results further show that models trained only on source-domain data struggle to generalize to the mountainous target domain because of large distributional discrepancies. Quantitative comparisons indicate that MR-AFDANet achieves the best overall performance， obtaining the highest IoU， F1-score， and Recall among all evaluated methods. These improvements confirm that combining multi-scale feature fusion with the SRADecoder enhances both global consistency and fine-grained boundary precision. Visual analysis further verifies this advantage： MR-AFDANet produces more continuous and complete road extractions under severe occlusion， maintains connectivity along steep and curved routes， and reduces false detections on spectrally similar surfaces.ConclusionWe present AsiaMountain-Road， the first high-resolution， multi-country dataset dedicated to road extraction in mountainous regions. Covering 29 representative areas across 17 Asian countries， it provides detailed road-mask annotations together with auxiliary contextual and topographic data， filling an important gap in current remote-sensing resources. Baseline experiments show that existing state-of-the-art methods struggle with narrow， winding， and occluded roads in complex terrain. To address these challenges， MR-AFDANet integrates SEAF modules， a dual-resolution feature-fusion mechanism， and the SRADecoder to recover thin road structures more precisely and improve robustness across highly variable mountain landscapes. Overall， AsiaMountain-Road narrows the gap between real-world geospatial challenges and current computer-vision capabilities by offering both a demanding benchmark for algorithmic innovation and a valuable resource for applied geospatial analysis. The dataset and code associated with this paper are available at https://cstr.cn/31253.11.sciencedb.j00240.00109 and https://github.com/wyl-ucas/AsianMountain-DataSet.关键词：Mountainous Road Extraction;high-resolution remote sensing;Remote Sensing Benchmark Dataset;Complex Terrain Road Detection;Deep Learning Semantic Segmentation28|39|0更新时间：2026-09-04 -
摘要：ObjectiveMotion blur is a common degradation in image acquisition that severely degrades three-dimensional scene reconstruction and novel view synthesis. In practice， motion blur arises from prolonged camera exposure or relative motion， causing loss of high-frequency details and geometric degradation. Existing deblurring methods based on neural radiance fields or 3D Gaussian Splatting have made notable progress. However， two critical limitations remain. First， camera trajectory modeling relies on discrete interpolation or standard neural ordinary differential equations， lacking physical consistency guarantees. Second， a single global transformation is adopted， which is insufficient to capture complex non-uniform motions such as curved trajectories and acceleration phases. This study aims to develop a motion deblurring 3DGS framework that ensures physically consistent trajectory modeling and enables precise representation of complex non-uniform motions.MethodThe proposed framework introduces two core technical contributions. First， a Hamiltonian dynamics framework based on symplectic geometry is established for continuous camera trajectory modeling. The camera state is represented as a position-momentum dual vector in phase space， coupled through the canonical symplectic structure. A symplectic neural derivative network parameterizes the Hamiltonian， and a symplectic Euler integrator performs numerical integration by alternately updating position and momentum components. This alternating update preserves the symplectic structure at the discrete level， thereby guaranteeing trajectory continuity and long-term numerical stability. The camera pose is decoded from the position component via the Rodrigues formula and SE（3） exponential map. Second， a dual-path adaptive motion decoupling framework is designed. This framework decomposes motion into two independent branches： a rigid motion branch that inherits the symplectic trajectory estimation to ensure physically consistent global camera motion， and a continuous motion refinement （CMR） branch that learns residual adjustments in an unconstrained space to compensate for discretized integration errors. An orthogonality regularization is applied to the CMR rotation matrix to maintain geometric plausibility， and the two branches are combined via matrix multiplication to produce final virtual camera poses. For deblurring rendering， a 3D frequency constraint filter is applied to Gaussian primitives to suppress aliasing artifacts. Multiple virtual cameras are sampled along the estimated trajectory， each rendering a sharp image via 3DGS. A pixel-level weighted fusion module and a blur-region mask network are introduced to adaptively synthesize the deblurred output. The experiments are conducted on the Deblur-NeRF synthetic dataset （5 Blender-rendered scenes with blur simulated in linear RGB space）， the Deblur-NeRF real-world dataset （10 scenes captured with a Canon EOS RP camera）， and the ExBluRF extreme blur dataset （8 scenes with challenging camera motions）. Camera poses and initial point clouds are recovered using COLMAP. Three evaluation metrics are adopted： PSNR， SSIM， and LPIPS. Ablation experiments are designed to independently verify the contributions of the symplectic geometry module and the dual-path decoupling framework.ResultThe experiments are conducted on the Deblur-NeRF synthetic dataset， Deblur-NeRF real-world dataset， and ExBluRF extreme blur dataset. On the synthetic dataset， the proposed method achieves 30.74 dB in PSNR， 0.901 in SSIM， and 0.073 in LPIPS， outperforming all compared methods in PSNR and LPIPS. The PSNR improvement over the second-best method SMURF reaches 0.12 dB， while the LPIPS improvement over the second-best method Deblurring 3DGS is 0.032. On the real-world dataset， the proposed method achieves optimal performance across all three metrics： 28.45 dB in PSNR， 0.853 in SSIM， and 0.084 in LPIPS， surpassing all compared methods. On the ExBluRF extreme blur dataset， the proposed method maintains robust performance with 30.74 dB in PSNR， 0.752 in SSIM， and 0.307 in LPIPS， where the PSNR advantage over the second-best method Deblur-NeRF （28.87 dB） reaches 1.87 dB. The experimental results demonstrate that the proposed method achieves optimal performance on all evaluation metrics across the real-world and ExBluRF datasets， and optimal PSNR and LPIPS on the synthetic dataset， with particularly pronounced advantages in complex non-uniform motion scenarios.ConclusionThe proposed framework successfully addresses the stated objectives by integrating symplectic geometry-based Hamiltonian dynamics with a dual-path adaptive motion decoupling strategy. The symplectic neural ordinary differential equation ensures physical consistency and numerical stability in trajectory estimation， while the dual-path decomposition enables accurate modeling of complex non-uniform motions. The method demonstrates superior reconstruction quality on standard and extreme blur benchmarks. The findings suggest that enforcing physical consistency in camera trajectory modeling substantially benefits motion deblurring in 3D scene reconstruction. However， the current method is designed for static scenes and has limited handling capability for dynamic foreground objects. Additionally， the high-dimensional symplectic neural derivative network introduces non-negligible computational overhead. Future work will focus on reducing computational complexity while preserving trajectory modeling accuracy， as well as extending the framework to dynamic scenes.关键词：3D Gaussian splatting;motion blur;symplectic geometry;motion decoupling;novel view synthesis45|61|0更新时间：2026-09-03 -
摘要：ObjectiveVisible-infrared person re-identification （VI-ReID） aims to retrieve images of the same person across visible and infrared modalities captured by different cameras， and has attracted considerable attention due to its potential applications in intelligent surveillance and all-weather pedestrian retrieval. However， VI-ReID remains challenging because of the large modality discrepancy caused by different imaging mechanisms. Visible images usually contain rich color and texture information， whereas infrared images mainly describe thermal radiation and structural characteristics， resulting in significant distribution differences between the two modalities. Recently， deep learning-based VI-ReID methods have achieved remarkable progress by learning modality-invariant feature representations with identity annotations. Nevertheless， acquiring large-scale identity annotations for visible and infrared images is expensive and time-consuming， which limits their application in practical scenarios. Therefore， unsupervised visible-infrared person re-identification （USVI-ReID） has gradually become an important research topic. Existing USVI-ReID methods generally rely on clustering algorithms to generate pseudo labels and employ contrastive learning frameworks to optimize feature representations. Although these methods have achieved promising performance， the quality of learned representations is still affected by insufficient hard sample mining strategies. Most existing methods select hard samples according to feature distances and fail to consider the different optimization roles of hard positive and hard negative samples. In contrastive learning， hard positives should maintain identity consistency while introducing challenging cross-modality variations to reduce intra-class discrepancy， whereas hard negatives should emphasize visually similar but identity-different samples to improve inter-class discrimination. Therefore， directly adopting a unified hard sample selection strategy may introduce noisy positive samples or ineffective negative samples， which restricts the optimization ability of contrastive learning. To address this problem， this study aims to propose a role-aware hard sample modeling framework for improving unsupervised cross-modality representation learning.MethodA role-aware hard sample modeling framework is proposed to construct hard positive and hard negative samples according to their different optimization objectives. First， shared pseudo labels are obtained through cross-modality clustering， and modality-specific identity centers are constructed for each identity cluster. Based on the feature shift between identity centers of different modalities， hard positive samples are generated along the cross-modality shift direction. Unlike conventional hard positive mining strategies that select samples with large distances from cluster centers， the proposed method introduces challenging modality variations while preserving identity consistency， thereby improving cross-modality feature alignment. Second， a hard negative modeling strategy based on local cluster relationships is proposed. Specifically， intra-identity features are divided into multiple sub-clusters， and neighboring clusters belonging to different identities are searched according to feature similarity. Hard negative samples are then generated along the direction of neighboring clusters. By introducing locally confusing identity information， the proposed strategy enables the model to focus on difficult inter-class relationships and learn more discriminative representations. Furthermore， inspired by curriculum learning， a progressive training strategy is introduced to gradually adjust the difficulty of generated hard samples. Instead of introducing excessively difficult samples at the beginning of training， the proposed strategy progressively increases the generation magnitude， allowing the model to learn from simple samples to challenging samples and improving optimization stability.ResultExtensive experiments are conducted on two widely used USVI-ReID benchmark datasets， SYSU-MM01 and RegDB. SYSU-MM01 contains visible and infrared images collected from multiple indoor and outdoor cameras， with significant variations in illumination， viewpoint， and modality distribution. RegDB provides paired visible and infrared images and is widely used for evaluating cross-modality retrieval performance. Rank-1 accuracy and mean average precision （mAP） are adopted as evaluation metrics. Experimental results demonstrate that the proposed method achieves superior performance compared with existing state-of-the-art USVI-ReID methods. On the SYSU-MM01 dataset under the all-search protocol， the proposed method improves Rank-1 accuracy and mAP by 5.5 and 4.1 percentage points compared with the best existing method， respectively. On the RegDB dataset under the visible-to-thermal retrieval setting， the proposed method achieves 93.0% Rank-1 accuracy and 86.2% mAP. Ablation studies further demonstrate the effectiveness of hard positive modeling， hard negative modeling， and progressive training strategy.ConclusionThe proposed method provides a new perspective for hard sample modeling in USVI-ReID by considering the different optimization roles of hard positive and hard negative samples. By constructing cross-modality shift-based hard positives and locally confusing hard negatives， the proposed framework effectively enhances cross-modality consistency and identity discrimination. Experimental results on two benchmark datasets verify the effectiveness of the proposed role-aware hard sample modeling strategy， providing a potential solution for robust unsupervised visible-infrared person re-identification.关键词：computer vision;person re-identification;visible-infrared;unsupervised learning;hard sample learning;contrastive learning29|51|0更新时间：2026-09-03 -
Image steganalysis network integrating equal-resolution frequency bands and contrastive clustering AI导读
摘要：ObjectiveSpatial adaptive steganographic algorithms usually embed secret messages in image regions with complex textures or edges to reduce visual distortion and statistical detectability. Therefore， the resulting embedding modifications are weak， sparse， and highly content-dependent. These subtle traces are easily masked by natural image textures， local intensity changes， and background noise， which makes reliable image steganalysis difficult， especially under low or moderate embedding payloads. Although deep-learning-based steganalysis methods have achieved better performance than handcrafted feature methods， existing models still have several limitations. Most of them rely mainly on single residual preprocessing， which may not fully exploit frequency-band information related to steganographic modifications. In addition， convolutional structures are effective for local texture extraction but have limited ability to model long-range contextual relationships. Moreover， cover and stego features are often highly overlapped in the feature space， making it difficult for cross-entropy loss alone to form a clear decision boundary. To address these problems， this paper proposes an image steganalysis network integrating equal-resolution frequency bands and contrastive clustering， termed CRF-CCNet.MethodsThe proposed CRF-CCNet is designed from three aspects： residual preprocessing， feature extraction， and discriminative optimization. In the preprocessing stage， an equal-resolution Laplacian frequency decomposition module is constructed to extract high-， middle-， and low-frequency residual components without spatial downsampling. This design preserves weak embedding traces while capturing complementary residual responses in different frequency bands. SRM high-pass filter banks are further applied to each band to enhance steganographic residual information. Considering that different frequency bands contribute unequally to detection， an adaptive attention fusion mechanism is introduced to dynamically adjust their weights. In the feature extraction stage， a Multi-Path Feature Enhancement Module is designed with three parallel branches. A lightweight Swin Transformer branch is used to model global contextual dependencies， a depthwise separable convolution branch is adopted to extract local texture features with relatively low computational cost， and a central difference convolution branch is introduced to enhance neighborhood intensity differences and edge-related perturbation responses. The outputs of the three branches are integrated through a gated fusion mechanism， enabling the network to combine global， local， and differential information. In the classification stage， a Multi-Scale Contrastive Clustering Loss is introduced. Dynamic class prototypes are maintained for cover and stego samples at different feature scales， reducing intra-class distances while enlarging inter-class margins. Meanwhile， a cross-scale consistency constraint is imposed on stego features to maintain structural coherence among different feature levels.ResultsExperiments were conducted on the BOSSBase 1.01 and BOWS2 datasets using three typical spatial adaptive steganographic algorithms， namely WOW， HILL， and S-UNIWARD. The proposed method was compared with representative steganalysis networks， including SRNet， LWENet， FA-CLNet， and HSMNet， under the same data setting. Under the BOSS+BOWS2 training setting with an embedding payload of 0.4 bpp， CRF-CCNet achieves detection accuracies of 94.81%， 89.55%， and 92.47% for WOW， HILL， and S-UNIWARD， respectively， outperforming HSMNet by 0.83， 1.55， and 1.15 percentage points. Compared with the lightweight steganalysis network LWENet， CRF-CCNet improves the accuracy on WOW at 0.4 bpp from 90.59% to 94.81%， with a gain of 4.22 percentage points. Meanwhile， the number of parameters is reduced from 0.82 M to 0.67 M， corresponding to a reduction of approximately 18.3%. Although the FLOPs increase slightly from 19.45 G to 20.25 G， the two models remain at the same computational scale， and CRF-CCNet reaches an inference speed of 82.4 FPS. Ablation experiments further verify the effectiveness of the equal-resolution multi-band preprocessing， the multi-path feature enhancement module， and the multi-scale contrastive clustering loss.ConclusionsCRF-CCNet improves the representation and discrimination of weak steganographic perturbations by combining equal-resolution multi-band residual modeling， multi-path feature enhancement， and multi-scale contrastive clustering optimization. Experimental results show that the proposed method achieves better detection accuracy for several spatial adaptive steganographic algorithms while maintaining a relatively small parameter size and high inference efficiency. Therefore， CRF-CCNet provides an effective lightweight framework for spatial image steganalysis.关键词：image steganalysis;spatial adaptive steganography;equal-resolution frequency decomposition;multi-path feature enhancement;contrastive clustering38|50|0更新时间：2026-09-03 -
A multi-color artificial illumination image dataset for nighttime melon segmentation MelonNightColor AI导读
摘要：ObjectiveNighttime melon recognition， localization， and harvesting in protected agriculture rely heavily on artificial illumination. Under different artificial illumination configurations， the apparent color， brightness， contrast， texture， and boundary visibility of melon fruits and stems may change， which can affect semantic segmentation performance. A model trained under one illumination configuration may therefore perform poorly when applied to images captured under another configuration. However， existing agricultural image datasets rarely provide spatially corresponding nighttime images of the same scene under multiple artificial illumination conditions. To address this gap， we constructed and released MelonNightColor， a multi-color artificial illumination image dataset for nighttime melon semantic segmentation. The dataset is designed to support baseline model evaluation， illumination-aware image analysis， and cross-configuration generalization research.MethodImages were collected from May 16 to June 4， 2025， in a melon cultivation facility in Huzhou， Zhejiang Province， China. The cultivar was the "Xiaobainiang" melon. All images were captured at night， with artificial illumination as the only light source in the acquisition area. A 12 W portable RGB LED lamp was placed beside the camera at an approximately fixed distance from the melon plants. The camera position， shooting distance， lamp position， and illumination direction were kept consistent throughout the acquisition process. Images were captured using a GM1910 camera in professional mode， with a fixed aperture of f/1.6 and a focal length of 27 mm （35-mm equivalent）. The flash was disabled， while exposure time， ISO sensitivity， white balance， and focus were automatically adjusted. Seven illumination configurations were generated by switching the red， green， and blue LED channels： red， blue， green， cyan， purple， yellow， and white. Red， blue， and green were obtained by activating one channel， whereas cyan， purple， and yellow were generated by simultaneously activating the green-blue， red-blue， and red-green channel combinations， respectively. White was defined as the simultaneous activation of all three channels. At each shooting location， the camera and scene were kept stationary while the seven configurations were applied sequentially. Consequently， each location produced a group of seven spatially corresponding images. The current release includes 5，474 pixel-level annotated images from 782 shooting locations. An additional 1，218 groups （8，526 images） are publicly available without annotations； their pixel-level annotations will be released incrementally in future updates. The original images are 24-bit sRGB JPG files with a resolution of 4，000 × 3，000 pixels. For the annotated portion， each image is paired with an 8-bit single-channel PNG mask of the same resolution. The annotation task includes three semantic classes： background， melon fruit， and melon stem. When fruits or stems were partially occluded， only their visible regions were annotated. Annotation quality was controlled through manual labeling， independent cross-checking， automated file-pair and size verification， legal pixel-value inspection， foreground-ratio analysis， scene-level split checking， and visual spot checks. The data were divided into training， validation， and test sets at a ratio of 7：2：1， with the shooting location used as the minimum splitting unit. Thus， the seven images from the same location were assigned to the same subset， preventing scene-level data leakage. MelonNightColor is organized into seven single-color subsets and four combined-color subsets. The latter were formed by merging single-color images at the data-organization level and do not represent additional physical illumination conditions or unique images. Nine representative semantic segmentation models were evaluated， including UNet， PSPNet， UNet 3+， SwinUNet， VM-UNet， DenseNet， DenseUNet， DeepLab， and H-vmunet. All models used the same data split and training protocols. Images and masks were resized to 224 × 224 pixels， and the models were trained using AdamW with a combined cross-entropy and Dice loss. No color jittering， hue adjustment， channel permutation， or synthetic recoloring was applied. In addition， UNet was used to construct a 7 × 7 cross-color generalization matrix， with training illumination configurations as rows and testing configurations as columns.ResultsThe baseline experiments confirmed that MelonNightColor supports the training and comparison of mainstream semantic segmentation models under nighttime artificial illumination. Averaged across the seven illumination configurations and nine baseline models， the IoUs of the background， melon fruit， and melon stem classes were 99.05%， 86.39%， and 38.84%， respectively. The stem was the most challenging class because it was thin， small， frequently occluded by leaves and vines， and often located near cultivation supports or shadowed regions. The average mIoU varied among illumination configurations. White illumination achieved the highest average mIoU of 78.17%， followed by yellow illumination at 77.94%. Red and blue illumination produced relatively low average mIoUs of 72.93% and 73.23%， respectively. UNet 3+， UNet， DenseUNet， and VM-UNet achieved comparatively strong results across most configurations. The average mIoUs of the red-blue， red-green， green-blue， and red-green-blue combined-color subsets were 73.30%， 73.90%， 74.52%， and 74.23%， respectively. These results describe the corresponding data organization and evaluation protocols rather than directly demonstrating the superiority of one physical light source over another. The cross-color generalization experiment revealed a substantial performance difference between matched and mismatched illumination conditions. UNet achieved an average mIoU of 80.38% when trained and tested under the same illumination configuration. When the training and testing configurations differed， the average mIoU decreased to 44.25%. In the most challenging example， training under yellow illumination and testing under red illumination produced an mIoU of 17.90%. However， several mismatched combinations retained relatively high performance， such as red-to-yellow， blue-to-white， and purple-to-white testing. These results indicate that the effect of illumination changes varies across color combinations and that MelonNightColor can support research on illumination robustness， color normalization， domain adaptation， and cross-condition semantic segmentation.ConclusionMelonNightColor provides a publicly available benchmark for nighttime melon semantic segmentation under multiple artificial illumination configurations. Its main contribution is the grouped acquisition design， in which each shooting location is represented under seven RGB channel combinations. The current annotated portion is accompanied by pixel-level annotations for background， fruit， and stem. The dataset provides predefined scene-level data splits， seven single-color subsets， four combined-color subsets， nine-model baseline results， and a UNet-based 7 × 7 cross-color generalization matrix. These resources enable researchers to study conventional segmentation accuracy as well as the effects of artificial illumination on image appearance， model stability， and cross-configuration transferability. The dataset was collected using a specific cultivar， camera， LED lamp， acquisition site， growth stage， and automatic exposure process. Therefore， the reported results should be interpreted as empirical benchmarks for the stated acquisition conditions rather than causal estimates of the independent effects of spectrum， illumination intensity， or camera parameters. MelonNightColor is publicly available through the Science Data Bank of the Chinese Academy of Sciences under the CC BY-NC-ND 4.0 license. Its DOI ishttps：//doi.org/10.57760/sciencedb.j00240.00237.关键词：facility agriculture;nighttime image;melon segmentation;artificial illumination;multi-color dataset;semantic segmentation31|49|0更新时间：2026-09-03 -
摘要：ObjectiveLiver cancer is one of the malignant tumors with the highest mortality worldwide. Accurate tumor segmentation based on computed tomography （CT） is of great significance for clinical diagnosis， surgical planning， and efficacy evaluation. Although deep learning has been applied in the field of medical image segmentation， under the fully supervised training paradigm， precise segmentation of liver tumors is constrained by the scarcity of high-quality annotated data. Specifically， the complex shapes and fuzzy boundaries of liver tumors， the high cost of manual labeling， and inter-observer variability lead to the scarcity of available high-quality training data， which has become a key bottleneck restricting the clinical application of such models. Although data synthesis provides a way to alleviate data scarcity， some existing methods may generate samples inconsistent with clinical anatomical rules. Furthermore， unnatural artifacts in synthetic images may cause models to learn shortcut cues， limiting their applicability to real tumor CT cases. To address these challenges， an annotation-free liver tumor segmentation framework， named Know2Seg， is developed. The objective of this framework is to alleviate the dependence of segmentation model training on real tumor annotations while making the synthesis process consistent with clinical rules and guiding the model to learn meaningful tumor-related features from synthetic data.MethodKnow2Seg mainly includes clinical knowledge-guided liver tumor synthesis and a training architecture. In the synthesis stage， rather than relying on generative networks， relevant clinical imaging knowledge is incorporated into the synthesis process in the form of computable constraints. The synthesis process contains four steps： tumor location sampling， morphological simulation， texture synthesis， and image fusion. These steps are designed to make the generated tumors more consistent with clinical and pathological characteristics. In location sampling， ellipsoid-constrained sampling and local anatomical checking are used to reduce anatomically unreasonable tumor placement. In morphological simulation， synthetic tumors are divided into different size categories， including micro， small， medium， and large lesions. Elastic deformation are used to generate irregular and anisotropic tumor shapes that conform to pathological features. In texture synthesis， random noise is processed by total variation denoising， wavelet decomposition， inverse reconstruction， and gaussian filtering to simulate the heterogeneous density patterns of liver tumors. In image fusion， a density attenuation field modulated by tumor texture is applied to healthy liver CT images. This strategy preserves high-density anatomical structures， such as vessels， and reduces unrealistic truncation artifacts caused by simple copy-paste operations. In the training stage， a training architecture informed by the idea of image disentanglement is used to support the model's generalization performance on real tumor CT images. Instead of directly learning the mapping from synthetic images to segmentation labels， the training objectives encourage the network to learn representations of healthy anatomy and lesion texture. The network outputs reconstruction maps and semantic segmentation logits. The segmentation logits are further converted into a tumor posterior probability map， which can be used as the blending weight for the reconstruction maps. Through this design， the model learns segmentation together with the reconstruction of healthy anatomy and tumor texture， which distinguishes normal structures from abnormal lesions. Four losses are jointly optimized， including healthy liver reconstruction loss， tumor texture reconstruction loss， segmentation loss， and blending consistency loss. With the intermediate ground truths generated during synthesis， this multi-task design makes the model learn more distinguishable representations of normal anatomy and abnormal lesions.ResultExperiments are conducted on public three-dimensional abdominal CT datasets. Know2Seg is compared with representative annotation-free tumor segmentation methods， including anomaly detection-based methods and synthesis-based methods. The results show that Know2Seg achieves promising segmentation performance. It obtains a dice similarity coefficient （DSC） of 62.9%， a normalized surface dice （NSD） of 65.5%， a surface distance （SD） of 14.0 mm， and a 95% hausdorff distance （HD95） of 40.4 mm. These results indicate that the framework performs favorably among the compared annotation-free methods. Since no real tumor annotations are used during training， the performance also suggests that the knowledge-guided tumor synthesis strategy can provide effective supervisory signals for downstream tumor segmentation， while the training architecture can improve the model's generalization ability on real data. The DSC and NSD results suggest good volumetric overlap and boundary consistency. The SD and HD95 results further indicate that the predicted tumor boundaries are relatively consistent with anatomical structures. Ablation experiments and hyperparameter sensitivity experiments further verify the effectiveness and robustness of the framework. These analyses indicate that both the synthesis strategy and the training architecture contribute to the overall performance.ConclusionBy transforming relevant clinical imaging knowledge into computable synthesis constraints and using a training strategy informed by image disentanglement to reduce the risk of shortcut learning， an annotation-free learning framework characterized by both interpretability and robustness is constructed. It does not require the pretraining of complex generative networks and helps reduce the reliance of model training on manual annotations.关键词：annotation-free learning;3D tumor segmentation;clinical knowledge;data synthesis;image disentanglement89|63|0更新时间：2026-08-24 -
摘要：ObjectiveExisting scene text editing methods mainly focus on monolingual editing within image patches containing single textual contents， and thus fail to support cross-lingual text editing. To address this limitation， this paper proposes SceneTrans， an end-to-end English-to-Chinese scene text localization-and-translation framework that enables accurate text localization in natural images， cross-lingual translation from English to Chinese， and visually consistent scene text editing.MethodUsing the multimodal large model MiniCPM-V 2.6 as the baseline， an fine-tuning strategy based on position-enhanced prompt templates is designed to jointly learn text localization and cross-lingual translation. A Chinese glyph structure encoder is introduced to explicitly model Chinese character shape priors， together with a glyph recognition supervision mechanism to ensure the structural accuracy of generated image text. To address the scarcity of cross-lingual scene text translation data， a dedicated English-Chinese paired dataset， SynthTrans， is constructed by collecting diverse English-Chinese compatible fonts and synthesizing 200，000 paired natural scene text images with complex glyph transformations such as curvature and inclination.ResultExperiments are conducted on public scene text detection benchmarks and the constructed SynthTrans dataset for comparison with existing methods. In the image text localization task， the proposed method achieves an F1 score comparable to mainstream detection models. In the image text editing task， an SSIM of 0.622 and a PSNR of 18.51 are obtained， while the text rendering accuracy （ACC） is improved to 71.13%. Overall， the framework demonstrates efficient and effective performance for cross-lingual scene image text translation.ConclusionThe proposed end-to-end seek-and-translation framework accomplishes an integrated “localization-translation-editing” pipeline for natural scene text images， generating Chinese scene text with accurate glyph structures and style consistency with the original text. It overcomes the limitations of previous methods that cannot perform localization or cross-lingual editing， offering a new technical solution and dataset foundation for cross-lingual scene text seek-and-translation in natural images.关键词：Scene Text Detection;Scene Text Editing;Text Image Localization and Translation;Large Multimodal Model Fine-tuning;diffusion generation64|71|0更新时间：2026-08-24 -
摘要：ObjectiveRecent advances in artificial intelligence generated image （AIGI） synthesis have improved the efficiency and quality of visual content creation， but they have also introduced information security risks. Realistic generated images can be misused for identity forgery， fake news fabrication， and evidence manipulation， making real-fake image discrimination an important task in digital media forensics. In image forgery detection， fine-grained semantic differences often exist between real and generated images in training and evaluation datasets. These differences can create spurious correlations between semantic content and authenticity labels， resulting in semantic bias. Such semantic bias can be regarded as a shortcut learning phenomenon： instead of learning forgery traces left by the generation process， detectors may rely on distinguishable semantic content differences between real and fake images for classification. This means that the high performance achieved by detectors on existing benchmarks with the help of semantic shortcuts may not fully reflect their generalization ability when semantic shortcuts are weakened. To address this problem， this paper proposes AIGI-Align， an instance-level semantic alignment dataset and benchmark for AIGI detection.MethodTo mitigate semantic bias from the data perspective， AIGI-Align is constructed under an instance-level semantic alignment constraint. The core idea is to use each real image as a semantic reference and generate corresponding fake images that are aligned with it in fine-grained semantic dimensions. Real images are selected from ImageNet and filtered according to image resolution to ensure basic visual quality. A multimodal large model is used to verify whether the image content matches the target category， ensuring semantic category correctness and reducing category mismatch. Unlike conventional datasets that rely on coarse category labels such as "a photo of ［class］" as generation prompts， this work introduces a vision-language model to generate fine-grained structured descriptions for each real image. These descriptions cover four visual dimensions： subject description， photographic style， image details， and background elements. These image-level descriptions are then used as text prompts to drive generative models， so that the generated images are constrained by the semantic content of the corresponding real images rather than by a coarse class label alone. Based on these prompts， 12 generative models are used to synthesize fake images， covering U-Net-based latent diffusion models， diffusion Transformers， rectified-flow Transformers， cascaded diffusion models， and a text-to-video model used in a single-frame setting. After generation， the fake images are checked by a multimodal large model for semantic category consistency， and samples with obvious visual artifacts are removed through manual inspection. The dataset contains 532 000 high-resolution images， including 70 000 real images and 462 000 generated images. It covers 14 semantic categories and 12 generative models， providing both instance-level semantic alignment training data and a controlled benchmark for evaluating cross-generator generalization.ResultExperiments show that AIGI-Align improves the degree of semantic alignment between real and fake images， alleviates the interference of semantic bias in the decision process of detectors， and improves their cross-generator generalization ability. Semantic bias analysis further indicates that existing detectors may use semantic differences between real and fake images for shortcut classification. We conduct a systematic evaluation on eight representative AIGI detectors， including CNN-based methods， frequency- or transformation-based methods， and methods based on pretrained visual representations. Evaluation is performed on ForenSynths， UniversalFakeDetect， GenImage， and the proposed AIGI-Align benchmark. Experimental results show that， under the conventional training setting， some detectors exhibit a performance drop in the semantic alignment evaluation scenario. This suggests that their original performance may be partially supported by semantic shortcuts rather than by robust recognition of forgery traces. When real and fake images are more consistent in semantic content， the semantic differences available to detectors are weakened， making the evaluation more focused on whether detectors can capture generation-related traces. After introducing instance-level semantic alignment training data， most detectors achieve improved cross-generator detection performance on both AIGI-Align and public benchmarks. These results indicate that instance-level semantic alignment data can reduce detectors' dependence on semantic content differences and encourage them to learn more general forgery-related cues.ConclusionThis paper presents AIGI-Align， a large-scale instance-level semantic alignment dataset for AIGI detection. In terms of training， instance-level semantic alignment data can guide detectors to focus more on generation-related forgery traces and reduce their reliance on semantic content differences during real-fake discrimination， thereby improving their generalization ability to unseen generative models. In terms of evaluation， the AIGI-Align test set provides a controlled scenario in which semantic variables are constrained， making it possible to evaluate the actual generalization performance of detectors when semantic shortcuts are weakened. Our dataset is available athttps：//huggingface.co/datasets/likeYousif617/AIGI-Align.关键词：artificial intelligence generated image;image forgery detection;semantic bias;semantic alignment;dataset;benchmark82|84|0更新时间：2026-08-24 -
Multimodal brain tumor segmentation network fusing explicit boundary prior and decoupled attention AI导读
摘要：ObjectiveBrain tumors， particularly gliomas， are among the most aggressive and heterogeneous malignancies in the central nervous system， with high morbidity and mortality. In clinical practice， precise and reliable segmentation of brain tumor sub-regions using multimodal magnetic resonance imaging （MRI） including native T1-weighted （T1）， post-contrast T1-weighted （T1ce）， T2-weighted （T2）， and fluid-attenuated inversion recovery （FLAIR） is a fundamental prerequisite for initial diagnosis， surgical planning， targeted radiotherapy， and postoperative survival prediction. However， manual delineation by radiologists is exceptionally time-consuming， highly subjective， and prone to intra- and inter-observer variability. While deep learning-based automated segmentation methods have made substantial progress， existing architectures frequently encounter bottlenecks. Conventional convolutional neural networks （CNNs） are constrained by their localized receptive fields， limiting their ability to model long-range semantic dependencies. Conversely， vision Transformers （ViTs） excel in capturing global contexts but suffer from patch-partitioning mechanisms that inherently disrupt pixel-level continuity， leading to the loss of high-frequency geometric details. Moreover， owing to the strong morphological heterogeneity of gliomas， and to the difficulty of boundary modeling and the insufficient fusion of spatial information with deep semantic features in existing networks， segmentation is prone to localization deviations， semantic confusion， and topological fractures in low-contrast regions.MethodTo systematically address these critical bottlenecks， this study proposes boundary prior enhanced network （BPE-Net）， which integrates explicit boundary priors and spatial-channel decoupled attention to couple explicit geometric perception with feature enhancement. The proposed BPE-Net is built upon an end-to-end 2D encoder-decoder architecture， comprehensively optimized to balance high-level semantic representation and low-level topological reconstruction. To explicitly capture high-frequency morphological details， we design an auxiliary-supervised boundary predictor module （BPM）. Unlike traditional edge-detection operators that are vulnerable to medical imaging noise， the BPM adaptively aggregates multi-scale hierarchical features （from the first three down-sampling stages of the encoder）. It projects these features into a unified resolution via bilinear interpolation and utilizes category-aware prediction heads to generate independent， three-channel spatial boundary probability maps for the whole tumor （WT）， tumor core （TC）， and the enhancing tumor （ET）. The BPM is supervised using category-specific， two-pixel-wide contour labels generated from the corresponding ground-truth masks. Specifically， the contour of each binary mask is extracted and rendered with a width of two pixels， providing explicit spatial supervision for category-aware boundary learning. Subsequently， to effectively integrate this geometric prior into the deep semantic space without introducing early-stage structural noise， we strategically deploy the edge feature enhancement module （EFM） exclusively at the high-resolution decoding stages （up3 and up4）. The EFM introduces a sophisticated spatial-channel decoupled attention mechanism. The spatial attention branch operates entirely on the boundary priors generated by the BPM to highlight high-frequency edge regions， while the channel attention branch dynamically recalibrates feature importance based on the decoder's intermediate representations using global average pooling and multi-layer perceptrons. The decoupled attention maps are then fused through element-wise multiplication and smoothly injected into the main network backbone via a residual connection parameterized by a learnable scaling factor. Furthermore， to rectify the inherent volumetric bias of standard regional losses （such as the traditional Dice loss）， which disproportionately favor large target areas （WT） while neglecting small， complex regions （ET）， we introduce a boundary weighted Dice loss （BWD）. The BWD incorporates a pixel-level spatial weight matrix derived from ground-truth boundaries. By assigning the highest penalty coefficients to the most ambiguous and microscopic edge pixels （especially for the ET region）， the BWD guides the network to focus more gradient updates on hard-to-classify transitional zones， thereby improving local discriminative capability.ResultExtensive and rigorous evaluations were conducted on two internationally recognized public benchmark datasets： BraTS 2019 （brain tumor segmentation 2019） and BraTS 2020， comprising diverse multi-institutional， multi-scanner MRI scans. The quantitative evaluation metrics included the Dice similarity coefficient （Dice） for regional overlap and the 95% Hausdorff distance （HD95） for boundary spatial distance. The proposed BPE-Net was compared with seven mainstream segmentation networks under identical data splits and training settings. On BraTS 2019， BPE-Net achieved an average Dice of 93.69% and an average HD95 of 1.96mm， exceeding the best-performing comparative method Swin-Unet （86.26%， 3.19mm） by 7.43 percentage points in average Dice and reducing the average HD95 by 1.23mm， with Dice gains of 8.12 and 9.91 percentage points for TC and ET， respectively. On BraTS 2020， BPE-Net achieved an average Dice of 92.51% and an average HD95 of 2.33mm， exceeding Swin-Unet （83.91%， 3.76mm） by 8.60 percentage points in average Dice and reducing the average HD95 by 1.43mm； in the ET region， its Dice reached 88.55%， 10.43 percentage points higher than the best comparative result. Qualitative visualizations further corroborated these findings， demonstrating that BPE-Net produces refined and topologically coherent segmentation masks that align well with expert annotations， reducing the over-segmentation and topological fractures frequently observed in baseline models. Comprehensive ablation studies verified the individual and synergistic contributions of the proposed components. The integration of the EFM alone improved the baseline average Dice from 86.31% to 93.57%， while the further addition of the BPM pushed the performance to 93.69%. Ablation on feature injection levels confirmed that the up3 and up4 combined strategy provides the optimal equilibrium between preserving deep semantics and recovering shallow structural fidelity.ConclusionIn summary， this study proposes an explicit boundary-guided network， BPE-Net， which helps mitigate the edge confusion issue in multimodal brain tumor segmentation. By innovating the auxiliary supervised extraction of category-aware boundary priors and executing precise feature injection via spatial-channel decoupled attention， the model achieves competitive and robust performance， particularly in delineating highly irregular and invasive tumor sub-regions. This proposed framework mitigates the volumetric bias inherent in conventional algorithms， providing an effective method for medical image segmentation and complex topology reconstruction. While the current architecture demonstrates superior accuracy， its computational complexity and reliance on complete four-modality MRI inputs represent certain limitations. Future research will focus on developing lightweight deployment schemes， exploring adaptive feature imputation techniques for missing-modality clinical scenarios， and validating the model's generalization capabilities across broader multi-center cohorts and other complex organ segmentation tasks. The source code of this paper is available at：https：//www.scidb.cn/s/eeiMBv.关键词：brain tumor segmentation;multimodal magnetic resonance imaging;boundary prior;auxiliary supervised learning;decoupled attention69|69|0更新时间：2026-08-24
- Postal code：100190
- Tel：010-58887035/58887030/58887418 Email：jig@aircas.ac.cn
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备05080539号-4
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0