最新刊期
卷 20 ， 期 11 ， 2015
- 摘要：The increasing large scale data puts forth a great challenge to multimedia computing. Different from traditional multimedia computing which is heavily based on hand-crafted features, deep learning (feature learning) recently achieves noticeable advance in multimedia computing. This paper presents the details of deep learning on multimedia retrieval and annotation, multi-modal semantic understanding as well as the video analysis and understanding, which tend to overcome the heterogeneity gap and semantic gap of multimedia computing in the setting of deep learning framework. On multimedia retrieval and annotation, deep learning-based "neural-codes" has been proposed and proves effective. Besides, deep learning is used for multi-modal semantic understanding to bridge the heterogeneity gap between different modals and the semantic gap between the bottom features and top semantic and deep learning-based compositional semantic learning is attracting increasing focus. Moreover, deep learning proves effective for video action recognition and for achieving a good representation of videos. However, the data-driven deep learning is easily affected by the noise in the data and is not ripe for online incremental learning. How to combine deep learning with crowdsourcing computing is a challenge and may be a future research direction. We analyze the existing methods of deep learning, and provide a new way to overcome the heterogeneity gap and semantic gap in deep learning framework.关键词：multimedia;large scale data;retrieval and annotaytion;semantic understanding;deep learning5053|384|5更新时间：2024-05-07
- 摘要：Establishing an accurate mathematical model and a design-effective algorithm is a dilemma in image restoration. The non-smooth energy functional model effectively describes image features but is difficult to use in a design-efficient computational algorithm. In this study, a new non-smooth energy functional regularization model that consists of fitting and regularization terms is developed. An efficient alternative iterative algorithm is deduced. First, for an image made blurry by system and salt-and-pepper noise in a tight frame domain, the fitting term is described by the L norm;the regularization term is established by the semi-norm of a weight-bound variation function. Second, the regularization model of image restoration is converted into an augmentation Lagrange model by introducing an auxiliary variable. Third, the transformed model is decomposed into two sub-problems by employing the variable splitting technique. Finally, by employing Fenchel transform and the fixed-point iterative principle, the sub-problem is transformed into dual and relaxed iterative sub-problems. The convergence property of the sub-problems is proven. An alternate iterative algorithm is proposed for the non-smooth property of the image restoration model. Compared with traditional algorithms, the proposed algorithm can effectively restore blurry images made so by system and salt-and-pepper noise and can increase the peak signal-to-noise ratio to approximately 0.5 dB to 1 dB. Results show that the proposed algorithm can effectively protect image edges and can achieve a high peak signal-to-noise ratio and structural similarity index measure. The proposed algorithm also has high convergence speed and can restore images rendered blurry by salt-and -pepper noise.关键词：regularization model;alternating iteration algorithm;image restoration;Lagrange multiplier2924|339|0更新时间：2024-05-07
- 摘要：Distortion in digital images is very common. To some extent, distorted image can affect research, such as analyzing and understanding images. In addition, the method of calculating the sharpness of an image is essential for the implementation of autofocus. Exploring its deeper mechanisms is of research importance. The performance of no-reference image quality assessment (IQA) has room for improvement. To upgrade the technology of sharpness assessment, an algorithm, which is called GI-F and is based on gradient information and human vision system (HVS) filter, is proposed. HVS is highly sensitive to gradient information. In the proposed algorithm, gradient information is first calculated using a gradient operator that researchers apply when computing image quality. Human studies in neurology also contributed to the development of other disciplines. Among visual cortex neurons, a mechanism occurs in which local excitation with higher amplitude can inhibit other impulses from global region. Therefore, HVS filter based on this characteristic is employed as a weighing function, in which the variable ranging from 0 to 1 stands for the relative impulse produced by each pixel in the picture, to obtain the sum of gradient information which represents the sharpness of image. Performance test can quantitatively evaluate different algorithms from the same perspectives, which generally include root-mean-squared error (RMSE), Pearson linear correlation coefficient (PCC), Spearman's rank-order correlation coefficient (SROCC), and time cost. To supplement, higher PCC and SROCC means that the score calculated using the proposed algorithm is more relevant with human vision system. Meanwhile, lower RMSE shows a smaller difference between two groups of samples. To ensure fairness of comparison among different algorithms, the test should be conducted under the same circumstance. The test is performed on public databases such as LIVE, TID2008, and CSIQ. Calculated results reveal that GI-F outperforms S3, CPBD and LPC-SI by using the metrics of RMSE, PCC and SROCC, which improved by 20.66%, 4.61% and 3.33%, respectively. In addition, the proposed method has lower computational complexity, and saves 79.72% computational time, compared with the currently best algorithm, BRISQUE. The proposed algorithm, applying gradient information and HVS filter, costs less time to compute objective sharpness, which is closer to the perceptive sharpness provided in public databases. In addition, such an algorithm can be widely used in sharpness assessment when reference images cannot be provided. In autofocus camera applications, more accurate and more stable sharpness results in improved performance of GI-F compared with other methods. Meanwhile, parallel computation is considered to save additional time as well.关键词：image quality assessment(IQA);no-reference sharpness assessment;gradient information;human vision system(HVS);high-pass filter(HPF);autofocus3448|275|3更新时间：2024-05-07
- 摘要：A foggy image has low contrast and low visibility. Traditional dehazing algorithms suffer from the halo phenomenon and color distortion in image dehazing. In consideration of the characteristics of the atmospheric veil, an image restoration algorithm based on relative total variation is developed in this study. Given that atmospheric scattering light does not affect texture information, the atmospheric veil is estimated accurately according to the main structure and texture information separated by relative total variation. An adaptive protection factor is then utilized to avoid distorting the restored image. Finally, the ideal result is obtained with a physical model, and brightness is adjusted by a curve. Compared with state-of-the-art dehazing methods, the proposed method can avoid halo artifacts or color distortion and can achieve a good dehazing result at distant scenes and in areas where depth changes abruptly. Experimental results show that the proposed method has robust scene adaptability and a fast computing rate because of the linear relation between time complexity and image size.关键词：relative total variation;atmospheric veil;dehazing;atmospheric scattering model5199|333|6更新时间：2024-05-07
- 摘要：To adapt to occlusion or other complex situations, an action recognition method is proposed which fuses multiple pose estimation features. Multiple pose features will be obtained using multiple action models. Each pose feature information includes key point positions and pose scores. Distinguishing key pointsare extracted from all train images and computing relative distances between point pairs. An action template is built using all features of the train images of the action. Multiple feature information consistent with multiple action templates are extracted from each test image from multiple pose features.Multiple features of the test image are-matched with the corresponding action template and then matched values are optimized using pose scores. The experimental results have shown that the average accuracy of the proposed method is approximately 2% better than some other state-of-the-art methods on VOC 2011-val set, and is approximately 6% better than some other state-of-the-art methods on Stanford 40 actions set. By fusing multiple pose features, the proposed method can adapt to occlusion and other complex situations and improve average recognition accuracy.关键词：action recognition;multiple pose estimations;template matching;occlusion2901|264|1更新时间：2024-05-07
- 摘要：With the high-speed development of computer technology and the need for video monitoring applications, face detection and tracking are gradually becoming a research hotspot and are widely used in public security, intelligent video surveillance, authentication, and others. Considering factors such as variation in lighting conditions, target obscuration, and long-term tracking, tracking human faces produces various errors that reduce the performance of the entire system. To resolve these problems, this paper presents an approach with a combination of detection and tracking technologies, and involves the three modules of detection, control and tracking (DCT). In the control modules, human face features are detected though haar features, and calculated with integral images. This reduces computational complexity and enhances system speed. The AdaBoost algorithm is then used to extract a face’s information in the detection module. Next, the information is transferred to the tracking module. The face target is tracked by using online random fern and Speeded Up Robust Features(SURF) algorithms in the tracking module. To improve tracking speed, 2-BitBP features are described in the random fern. Then, the P-N learning method is used to quickly locate similar objects. Similar targets and matching confidence of the target are determined based on the NCC algorithm. The similar target with the largest confidence value is considered the tracking target. Whereas, the control module is used to judge accuracy of the current tracking after each target is detected. The interactive unit in the control module transfers the testing results into the tracking modular for subsequent tracking. Three filters in the check unit are designed to filter check results in terms of scale size, target position, and similarity. This balances the performance of detection module and tracking module and avoids tracking failure because running the system for a long time and interference of similar targets. By using international standard data sets, the experimental results compared with LBP + Camshift + Kalman filter algorithm, the SEMI-supervised learning algorithm, and the Tracking-Learning-Detection (TLD) algorithm, show that the DCT approach proposed in this paper has good tracking effect on large scale target changes, occlusion, rotation, deformation and illumination changes. This effect can be achieved at the level with identification accuracy of over 95%, the average error recognition rate of 0.86% and missing recognition rate of 0.78%. On tracking accuracy, when light varies greatly, the experiments of the target center offset show that the DCT approach can still track the target position accurately. In terms of tracking performance, the average running time of this approach is 46.25 ms, and its effective rate is 98.25%. DCT can eliminate error accumulation, and automatically detect the target once tracking fails. Moreover, it can reduce the influence of environment lighting, obscuration, and affine transformation and meet performance requirements of real-time tracking in the augmented reality system. This approach can improve real-time performance and robustness of the system when applied to a video face tracking system.关键词：face detection;tracking;control;AdaBoost;random fern3285|303|3更新时间：2024-05-07
- 摘要：Several low-rank matrix decomposition-based approaches have been proposed for moving object detection in recent years. However, most of these methods use the nuclear norm to substitute rank functions for optimization. As a result, the precision of background recovery is relatively low. Another problem is the failure of these methods to use prior knowledge of the regional continuity of foreground objects, which is important information for object detection. To solve these issues, we propose a novel object detection method that combines the weighted non-convex nuclear norm and the regional continuity of the foreground object. The new object detection model is designed on the basis of the robust principal component analysis. The proposed model uses the weighted non-convex nuclear norm to replace the traditional nuclear norm for low-rank constraints. Furthermore, the prior knowledge of the regional continuity of the foreground object is added to restrain the clustered objects. By using this model, the recovered low-rank matrix becomes the background image matrix, and the large sparse noise matrix becomes the foreground object matrix. Experiments demonstrate that the proposed method outperforms other low-rank decomposition-based approaches in both the simulated data and real sequences. Specifically, the proposed methodology shows an increased projected target detection performance that is 25% and 2% greater than that of RPCA and DECOLOR. With respect to the two approaches, the proposed method reduces background recovery errors by about 0.5 and 0.2. The non-convex relaxation of rank functions possesses better properties than the convex one in approximating matrix ranks, which is useful in restoring background images in motion object detection. The regional continuity of foreground objects allows the efficient exclusion of scattered outliers to enhance the effect of the objects detected. Therefore, this method can detect moving targets accurately in complex scenes, such as those with dynamic backgrounds and illumination-changing scenarios. However, the proposed method is not ideal for multi-object detection in small areas because of regional continuity requirements.关键词：moving object detection;low-rank matrix decomposition;weighted nonconvex nuclear norm;regional continuous;matrix recovery3476|277|4更新时间：2024-05-07
- 摘要：Moving object detection and tracking is an important step for many computer vision applications. Considering the presence of background movement and target motion, moving object detection and tracking in dynamic background became more complex. Thus, such detection and tracking is one of the important, difficult problems in machine vision research. This article proposed a method based on the OPTICS clustering algorithm and the model of probability area, aiming to increase target-tracking performance and accuracy of a moving camera. This article combined the advantages of SIFT feature description and Harris corner detection, aiming to solve the problem of large computation of SIFT algorithm and large error of object area. First, the Harris-Sift feature points detection are introduced. The feature points are not only the significant corner points, but also possess scale invariant features. After establishing the key points of feature description, adjacent feature point is matched by Best Bin First algorithm, improving tracing accuracy and robustness of feature point. To analyze the motion of the feature point better, the motion vector of the feature points is converted into the corresponding optical flow coordinate, adopting gird, and data binning technique to determine the scope of the search range of data points. Based on to the difference in moving target and background motion vector, the improved OPTICS algorithm is introduced to cluster on the grid structure, which could significantly reduce computation time of the algorithm. After obtaining the estimation of the moving object area, this article defined a class of feature points that are the most widely distributed in the image as the background, whereas other feature points represent a moving object. After separating the background and moving objects to track the target continuously, the tracking strategy is to search for the upper frame moving target area's feature points in the next frame. Although the OPTICS algorithm removed most of the noise, there are still errors between the real region of the objects and the estimation of the region. Next, based on each moving target, an independent probability model is built. This article establishes an independent object probability model based on continuous video frames for each moving object, along with the iteration of testing area, the real aiming area is then extracted. In this article, a new method is proposed to solve the problem of low accuracy and tracking performance of multiple moving objects tracking in a complex environment. The experimental results demonstrate that the proposed algorithm, which is based on the premise of meeting real-time, can extract the multiple moving objects from the complex moving background accurately. Experimental results show that the proposed Harris-Sift feature point detection algorithm can greatly reduce calculation time. Thus, the algorithm can meet real-time requirements. The proposed method can accurately separate each moving object from the background not only indoors but also under complex outdoor environments;it can also adapt to changing illumination and camera movement.关键词：computer vision;image processing;multiple moving object tracking;moving background;clustering3008|349|2更新时间：2024-05-07
- 摘要：To simulate human cognitive process of repeated intercomparison and deliberate refinement, an image pattern cognition method and its calculative model with a simulated feedback mechanism are explored. The feature space is progressively optimized under the different cognitive demands via evaluative index system of cognitive result analysis. Layered feedback cognitive mode from global to local is realized. The model is successfully applied to burning state cognition of rotary kiln flame image. First, a simulated feedback mechanism-based image cognitive model with interactive running of training layer and cognitive layer is proposed;and the objection, structure and function of the model are elaborated. Second, analysis and evaluation of cognitive result-based simulated feedback mechanism is designed, and the evaluative index system of cognitive result analysis is constructed. Third, specific to the application of rotary kiln flame image-based burning state cognition, the original flame image feature space is constructed using KPCA to establish image pattern cognitive information system in the sense of maximum cognitive information content. Then, the compressed cognitive information system is established based on attribute core calculation and Mahalanobis metric function. Finally, under different cognitive demands, cognitive result is evaluated based on evaluative index system of cognitive result analysis to update the compressed cognitive information system, and simulated feedback intelligent cognition of burning state is realized. The simulation experimental study is carried out according to actual acquisition of collected rotary kiln flame image, and the average cognitive accuracy is 91.78%. Experiment results show that, relative to the existing open-loop method, our method cognates repeatedly similar samples which are hard to distinguish through the simulated feedback mechanism. Thus, cognitive accuracy is improved effectively.关键词：image pattern cognition;α-entropy;simulated feedback mechanism;analysis and evaluation of cognitive result;burning state cognition3092|266|2更新时间：2024-05-07
- 摘要：Spline curve surface is an important part of CAGD. To obtain the parameter spline curve on the hyperbolic paraboloid, the four rational basis functions presented in this paper have a changeable control handle and a unary function. A hyperbolic paraboloid is constructed in the standard tetrahedron. On the surface, the base function defines a parameterized method with the parameters of the shape parameter spline curve surface. The basis function of spline through the rational parameters of the hyperbolic paraboloid is limited. A single-parameter rational basis function of spline is generated. Additionally, the nature of shape preserving and endpoints are considered. The nature of shape preserving and endpoints are considered. The curvature of two endpoints are derived as zero for a range of points. Consequently, a -continuity curve can be attached by this spline curve without any additional requirements. Experiment shows that we use this spline to raise the surface for spatial mesh, which also can implement -continuity with much freedom.关键词：basis function;rational spline;approximate spline;-continuity2507|254|3更新时间：2024-05-07
- 摘要：Existing structure-analysis-based algorithms of building detection in high-resolution synthetic aperture radar (SAR) images only consider straight-line and L-shaped buildings and utilize the shadowed areas of the highlighted lines on buildings in the detection process. In complicated scenes, shadowed areas cannot be accurately detected by restricting dark backgrounds and dense buildings;this condition results in large errors and low accuracy in building detection. Focusing on these problems, an algorithm based on the morphological hierarchical analysis of unsupervised building detection in high-resolution SAR images is developed in this study. The method is applied to single-polarization high-resolution SAR images. First, an improved alternating sequential filter (ASF) (i.e., extended ASF) is utilized to reduce the inherent speckle noise and eliminate the interference of background noise in the homogeneous regions significantly. Second, the differential morphological attribute profiles are calculated to implement the geometric structure feature extraction of buildings in a SAR image in a complex scene. Finally, post-processing methods, such as feature fusion and threshold segmentation, are performed to obtain intricate building information. Compared with other structure-analysis-based algorithms, the proposed method exhibits a higher detection rate and a smaller error for an urban area with high-density buildings. The precision and recall rates of the proposed method are 95.38% and 86.31%, respectively. The method also results in reduced false-alarm rates. The improvement and combination of ASFs and morphological attribute profiles are suitable for the detection of high-density buildings with different directions, sizes, and shapes.关键词：building detection;hierarchical analysis;morphological attribute profiles;denoising;high-resolution SAR3410|382|0更新时间：2024-05-07
- 摘要：When applying image classification algorithms on high-resolution images to extract roads, non-road areas do exist in the binary result. Meanwhile, the achieved roads are planar, which cannot be used directly for production and research purposes. In this case, a novel algorithm named neighborhood centroid voting is proposed to extract road centerlines. First, a neighborhood polygon for each road pixel is built by detecting the connective distance in each direction. Then, centroids of these polygons are voted for to extract road centerlines. At the same time, road width is estimated and the number of those directions, comparatively long connective distance is recorded to exclude non-road areas. Finally, morphological methods are applied to obtain thinned centerlines. A comparison is made between this algorithm and a reference method proposed by Zhang and Couloigner via experiments on a test image and two classified high-resolution aerial images with different road distributions. Results suggest that the quality of this algorithm for the respective two images is 80.6% and 79.0%. Taking less than 20% of the time of the reference method for dealing with actual images, this algorithm has a strong advantage because of its effectiveness. Additionally, this algorithm is more stable and can adapt to roads with varying widths. The proposed algorithm named neighborhood centroid voting is a centerline extraction algorithm capable of doing work corresponding to road refinement and centerline extraction in a conventional approach at the same time. Experimental findings suggest that this algorithm can detect roads effectively according to shape features, with resistance to disturbances, applicable to high-resolution classified images with roads and non-road areas mixed toge-ther.关键词：road extraction;centerline extraction;neighborhood centroid voting;shape feature;connective distance3586|322|3更新时间：2024-05-07
- 摘要：Spectral unmixing is a key hyperspectral remote sensing image processing technique. Many spectral unmixing algorithms have been proposed. Most of these algorithms are based on the assumption that pure pixels exist in the hyperspectral imagery. However, when pure pixels are lacking, the performance of these algorithms may deteriorate. The simplex volume minimization (VolMin) method provides a good means to overcome this problem. However, VolMin is a complex constraint optimization problem. Owing to the uncertainty that noise exists in an image, the algorithm is easily trapped into local optima. Thus, we introduce a swarm intelligence technique, i.e., differential evolution (DE) algorithm, into the VolMin procedure. By utilizing the powerful global searching capability and high-dimensional adaptability of DE, we develop a minimum simplex volume DE (VolMin-DE) spectral unmixing algorithm through VolMin-DE encoding. Synthetic mixture and real image data are utilized for comparative experiments on unmixing accuracy. The proposed VolMin-DE algorithm, vertex component analysis, minimum distance constrained non-negative matrix factorization (NMF), and minimum volume constrained NMF are compared. Experimental results indicate that the proposed VolMin-DE algorithm outperforms the other algorithms. The precision (spectral angle distance) of VolMin-DE can be increased by 7.8%, especially when the accuracy of 10 endmembers is improved by 41.3%. In the noise range of 20 dB to 50 dB, the performance of VolMin-DE varies from 1.9 to 3.2 (unit: degrees), whereas that of the traditional algorithm varies from 2.2 to 3.5. This result demonstrates that the proposed algorithm has better noise robustness than other algorithms. The proposed VolMin-DE algorithm can be applied to hyperspectral imagery regardless of the pure pixel assumption (maximum purity level equal to or greater than 0.8 is recommended). The algorithm does not require dimensional reduction;hence, error accumulation is avoided. VolMin-DE has good potential applications in hyperspectral unmixing.关键词：hyperspectral remote sensing(HRS);hyperspectral unmixing(HU);endmember;non-negative matrix factorization(NMF);differential evolution(DE)3584|244|2更新时间：2024-05-07
- 摘要：At present, the total number of the visually handicapped in the world is large and continues to grow. However, most seeing-eye robots are limited to the specific environment or single spaces and have difficulty adapting to changes in environment. Thus, these robots lack practical value to meet the needs of society. To overcome the space limitation and guide the blind in different environments, this paper proposes an effective method for guiding the blind based on the creation of environmental maps to expand the scope of the environmental adaptation of the seeing-eye robot. First of all, simultaneous localization and mapping (SLAM) is used to create a two-dimensional map of the environment, in which the seeing-eye robot localize itself and simultaneously build an environmental map based on observational features and position estimation;Then, the A heuristic search algorithm is used to plan a global path to find an optimal path without collision in the static map that has been created. The artificial potential field algorithm is then combined to avoid the dynamic obstacles detected by the seeing-eye robot when walking along the global path;Finally, the software control system of the seeing-eye robot is constructed using the robot operating system (ROS) framework so that the functional nodes can communicate with each other to obtain the required data based on certain rules, thereby making the whole control system more orderly and efficient. The experiments were carried out in three typical surroundings, such as an office, a long corridor of big building, and an outdoor rest area. The result of the experiments proved that, compared with other blind methods, the proposed method is applicable to a wider range and more flexible environment and is no longer limited to a single space or specific environment. In the process of creating the map, the feature estimated error is only within the range of 5cm-35cm when the total number of features is 30 and the position estimated error is less than 3m when the walking steps are up to 12000 steps. In planning the path, trajectory error is under 0.4m when the length is up to 120m. By comparison, the proposed method is more practical and superior for guiding the blind. This paper presents a method for guiding the blind applicable to various environments. The corresponding experimental results demonstrate that the map created by the method is consistent with the actual scene and the trajectory of the robot is essentially identical to the planned path in various environments, so its precision is relatively high. The method is generally applicable to indoor and outdoor areas of the daily activities of the visually handicapped and the method is sufficiently flexible to adapt to transformation in the environment. Therefore the method proposed by this paper is more practical and effective for guiding the blind.关键词：seeing-eye robot;various environments;mapping;path planning2496|539|0更新时间：2024-05-07
- 摘要：Spatial location retrieval is one of important pretreatment steps in remote sensing data searching. In order to improve the original Ray-algorithm performance and reduce the false alarm rate in retrieving catalogue data from massive remote sensing images, a new hybrid parallel strategy has been proposed based on MPI and OpenMP to implement the Ray-algorithm. First, two improved processing methods have been added to the classical Ray-algorithm to deal with two special cases between the point on the edge of the polygon and the ray intersecting on the endpoint. And then, we use multi-process based on MPI to do parallel computing on a PC cluster in program procedure and multi-threads based on OpenMP to do parallel algorithm implementation in which many threads are started synchronously to deal with each point of the polygon which is in another point or not. The experimental results show that the whole system will reach better speedup when the sum of started threads in all nodes is equal to the optimal threads number of the main node.Compared with serial algorithm, the retrieval time of hybrid parallel algorithm reduced by more than 50% and this algorithm is more efficient. The hybrid parallel strategy based on MPI and OpenMP is better than common serial execution and parallel execution based on MPI and OpenMP respectively. The hybrid parallel mode is generally applicable to the cluster environment parallel programs and it can used to deal with other images processing algorithms with parallel features.关键词：remote sensing;location retrieval;ray-algorithm;hybrid parallel strategy;optimal threads3053|316|0更新时间：2024-05-07
- 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