最新刊期
卷 21 ， 期 7 ， 2016
- 摘要：A random measurement matrix plays a critical role for the successful use of compressive sensing (CS) theory and has been widely applied in CS. However, a random measurement matrix requires a large storage space, which is unsuitable for large-scale applications. To reduce the storage space for a random measurement matrix, a method for CS signal reconstruction was proposed based on theory of semi-tensor product. We constructed a random measurement matrix, with a dimension smaller than and , where is the length of the sampling vector and is the length of the signal that we intend to reconstruct. Then, we used the iteratively reweighted least square reconstruction algorithm to obtain the estimated values of sparse coefficients. Experiments were conducted using column sparse signals and images with various sizes. During the experiments, the probability of exact reconstruction, error, and peak signal-to-noise ratio (PSNR), of the proposed method were compared with measurement matrices with different dimensions. The proposed algorithm outperformed a smaller storage space with a suitable PSNR performance. In this study, we proposed a new method to reduce the storage space of the measurement matrix for CS. The experimental results showed that if we appropriately reduced the dimension of the measurement matrix, then nearly no decline in the PSNR of the reconstruction was observed, but the storage space of the measurement matrix could be reduced by at least 1/4 or 1/16 times. All the results verified the validity of the proposed approach and demonstrated the significant potential for hardware implementation of the proposed sensing framework.关键词：compressive sensing;random measurement matrix;storage space;semi-tensor product;iteratively re-weighted;minimization2875|280|1更新时间：2024-05-07
- 摘要：The new generation 3D video adopts multi-view video plus depth as the main format. To improve the coding performance and quality of synthesized views, several new techniques, including depth modeling mode, segment-wise DC coding, and skip intra mode, are introduced into the current 3D-HEVC test model for depth intra coding. However, these tools dramatically increase computational complexity, which obstructs the practical applications for the 3D video. Thus, in this study, a fast depth intra coding algorithm is proposed to alleviate the burden of the encoder. Considering that 3D-HEVC reuses the well-known hierarchical coding structure, the mode information of the parent coding units (CU) can be reused to predict the most probable modes for the current CU. Consequently, only a subset of the intra mode candidates is employed to calculate the rough rate-distortion cost. According to rough modes and their rough costs by rough mode decision, the current CU will be classified into three categories, namely, smooth, directional, and edge. For each class, a different set of rough mode candidates will be selected. Simulation results show that the proposed algorithm achieves a reduction of 44% encoding time for depth intra coding while maintaining the quality of synthesized views, compared with the original test model HTM-13.0. Based on the mode information of parent CU and rough mode decision, the proposed method can significantly reduce the number of intra mode candidates. Consequently, the proposed method can provide a remarkable time reduction of depth intra coding with a negligible BD rate increase.关键词：three-dimension high efficiency video coding(3D-HEVC);depth map;intra coding;hierarchical coding structure;rough mode decision2829|280|2更新时间：2024-05-07
-
Fusion method via KNN with weight adjustment for the classification of image-text co-occurrence data
摘要：Most existing fusion methods are suitable in situations where the image classification method is identical with the text classification method. However, better classification methods of image and text for image-text co-occurrence data are not identical in many application scenarios. The decision benchmarks of different classification methods are not unified, which would reduce the classification precision of the fusion method. The fusion method based on KNN with weight adjustment for the classification of image-text co-occurrence data is proposed to overcome the problem. First, the softmax and multiple classification SVM are used to classify the image and text. The image's KNN model and the text's KNN model are constructed using the weighted classification decision values of image and text instances, which are correctly discriminated on a training dataset. Then, the classification decision values of the test instance are predicted by the image's KNN model and the text's KNN model. The image classification probability and the text classification probability of the test instance are determined by the number of each category in the nearest neighbors, and the classification probabilities would unify the classification decision benchmarks. Finally, the fusion coefficient is calculated by the number of image instances and the number of text instances discriminated correctly on a training dataset applied to fuse the classification probabilities of the image and text for the test instance. We performed an experiment on the Attribute Discovery dataset and compared the proposed fusion method with the baseline method. Experimental results show that the proposed fusion method achieved higher classification precision than the image classification method and text classification method, and the proposed fusion method increased the average classification precision by 4.45%. Moreover, the proposed fusion method increased the average information integration ability for image-text co-occurrence data by 4.19%. The proposed fusion method unifies the classification decision benchmarks of the different classification methods for image and text and implements the effective fusion of image-text co-occurrence data. Therefore, the proposed fusion method has a better ability to integrate information, with a better classification performance.2926|266|2更新时间：2024-05-07 - 摘要：Face image beautification is a widely used technique in multimedia, such as digital cameras, mobile terminals, advertising, and video conferences. However, several defects of face image beautification algorithms still exist today, such as the bad effects on detail-rich eyes and hair areas, as well as the poor visual effects on overall facial beautification images. A fast facial beautification method based on skin color segmentation and smoother face image is proposed in this study to overcome the defects of face beautification. First, based on characteristics of the skin's smoothness affecting facial attractiveness, facial blemishes are removed. The biexponential edge-preserving smoother is used to smooth the face image for facial defects and keep the important information of image edge. Second, some background information will be lost during the process of achieving skin smoothness, thereby making the separation of the skin and non-skin area a highly significant process. Then, based on the adaptive hue histogram, rapid detection, correction, and segmentation to the color region are achieved. Third, the mask has been obtained using the Gaussian fitting, which can have a box blur for many times for skin feather. Then, the skin feather is used to fuse the smoothed image and the original image to preserve the details of the face image, such as hair and background. Finally, based on the required white and natural skin as the beauty standard of a portrait, face image brightness is speedily adjusted, other important details are enhanced by fitting the log curve, and facial beautification can be achieved quickly. Compared with different methods of face image beautification, we performed qualitative and quantitative evaluation on the same face. After comparing with other facial beautification algorithms, we determined that the proposed algorithm in this study is more effective on the smoothness of edge-skin blemishes and has better ability to beautify the face image in terms of edge preservation; in terms of time complexity, its computing rate is 12 times as fast as the previous algorithms, and it ensures faster beautification of the face image relative to the previous images; in terms of operating methods, the proposed algorithm is simple, considering it only uses an adaptive histogram of chromatic channel to correct and segment skin color region, and it can achieve significant effect in color and non-color regions; and edge connection is more natural. We have performed a user study for the effectiveness of a proposed algorithm, beautifying a large number of face images with different genders, ages, poses, and backgrounds from the Internet. Experimental results show that the proposed algorithm has strong adaptive capacity because it can perfectly remove facial defects of most face images and it can simultaneously keep the background information unchanged, making the skin white and natural. The effectiveness of overall beautification is significant. In general, with its good beautification effect, the proposed algorithm obtains consistent high praise from many users. Specifically, it can moderately smooth the edges of the areas with many details, leaving no traces of an artificial process. The proposed algorithm with fast beautification of face image has a wide range of practicality.关键词：image beautification;bi-exponential edge-preserving smoother;color histogram;skin-color segmentation;facial defects4175|438|1更新时间：2024-05-07
- 摘要：Face detection is constantly an active research subject in computer vision and pattern recognition. Face detection is also a constituent part of pattern recognition, artificial intelligence, information security, and many other disciplines. With video network coverage widely increasing in recent years, face detection has been increasingly used in the field of video surveillance. However, many factors require consideration in face detection, such as the complex environments, multiple faces, and face rotation angles. In view of these interference problems in nonideal condition, a cascaded neuron network based on a multi-layer probability state-restricted Boltzmann machine (P-RBM) is proposed in this study to overcome the challenge of accurately and rapidly detecting faces. The neurons of RBM only have two states, namely, activated and nonactivated; this state mode can inhibit the interference in the learning result induced by the inadequate active information, while it simultaneously increases the likelihood that the learning network falls into a local optimum caused by the shielding of relatively weak information. To solve this contradiction, the proposed method uses the probability state of neurons in RBM as their activation degree, which better models the activity state's continuous distribution of the neurons in the human brain. Using the probability state not only retains the weak active information but further decreases the effect caused by the former layer's miscalculation. Simultaneously, this method simulates the hierarchical learning mode in the human brain by cascading multiple P-RBMs. This cascaded network can achieve multi-layer nonlinear mapping and obtain the semantic feature of the input date by extracting the input data's separate level features. Furthermore, this cascaded network can learn the relationship hiding within the data to make the learned features be more promotional and expressive. Simultaneously, the number of the hidden layer's neurons decreases layer-by-layer to control the network's scale and enhance the robustness. Finally, the proposed method uses the layered training and the entire optimization to balance robustness and accuracy. The greedy layer-wise learning is used in the layered training to avoid the training error transferring in layers, thereby solving the problem of the multi-layer network easily falling into the local optimum. Furthermore, a preprocessing layer is used to detect the skin color area to reduce the number of neurons in the detection network and speed up the detection speed. Testing the single face detection performance in the LFW and FERET, the proposed method nearly achieves entirely accurate detection. Testing the video face detection in the PKU-SVD-B database, the missing detection rate and the false detection rate of the proposed method are all lower than that of the state-of-the-art methods, such as Adaboost and Adaboost combined with skin color detection, and its detection speed is faster. Moreover, the proposed method has a good detection performance for the face with a large rotation, which is tested in the CAS-PEAL database. Experimental results show that regardless of whether a static single face or video multi-face detection occurs under complicated conditions, apart from the faster detection speed and robustness against face rotation, the proposed method possesses lower false detection rate and lower missing detection rate. Aiming at the multi-face detection based on skin color, this method can significantly reduce the false detection rate.关键词：face detection;restricted Boltzmann machine(RBM);probability state-restricted Boltzmann machine(P-RBM);neural network3011|306|3更新时间：2024-05-07
- 摘要：Active contour models (ACMs) are efficient frameworks for image segmentation. They can provide smooth and closed contours to recover object boundaries with sub-pixel accuracy. Region-based ACMs identify each region of interest by using region and statistical information as constraints to guide the motion of the active contour. The most popular region-based ACM is the Chan-Vese (C-V) model, which has been successfully used in binary phase segmentation with the assumption that image intensities are homogeneous in each region. However, typical region-based models do not work well on images with intensity inhomogeneity. This paper presents a new level-set-based K-means ACM, which can effectively segment images with intensity inhomogeneity. The model is derived from a linear level-set-based K-means model, which is established based on the properties of the corresponding Euler-Lagrange equation of the traditional ACM. The most widely used region-based ACMs are used, namely, the global region-based ACM, C-V model; the local region-based ACM, the local binary fitting (LBF) model; the local image fitting (LIF) model; and the local correntropy-based K-means (LCK) model. Their fitting terms correspond to the classical K-means. The C-V model has a global segmentation capacity, that is, it can segment all objects in an image; however, it cannot handle images with intensity inhomogeneity and various noises. The LBF and LIF models possess a local segmentation property; thus, they can only segment the desired object with a proper initial contour. In this paper, we propose a new region-based ACM, which integrates the advantages of global and local region-based ACMs. The global information and local information are combined to avoid entrapment in the local minima, and the local correntropy-based adaptive weights are used to ensure robustness against noise and fast convergence. The proposed model can successfully detect objects in a noisy synthetic image with intensity inhomogeneity. Results of the experiments on medical images show that compared with the background models, the proposed model can yield competitive results. Furthermore, when different initial contours are used, the proposed model can still realize correct segmentation for inhomogeneous images, whereas the other models are easily trapped in the local minima. This segmentation results demonstrate that the proposed model is not only capable of obtaining better segmentation results but also robust against noise and initializations. A novel and robust ACM based on K-means is proposed to segment images with intensity inhomogeneity. Relying on the correntropy-based image features, the model uses local adaptive weights to withstand various noises. Moreover, the combination of local and global region information prevents the proposed model from being trapped into a local minimum. To avoid re-initialization and shorten the computational time, we use a signed distance function to regularize the level set function and adopt an iteratively re-weighted method to enhance the speed of our algorithm during the contour evolution. The experimental results show that our algorithm can achieve robust segmentation even in the presence of the intensity inhomogeneity, noise, and blur.关键词：image segmentation;active contours;level set method;K-means;intensity inhomogeneity;correntropy2835|311|8更新时间：2024-05-07
-
The polynomial potential function with a parameter and transition curve based on Metaball technology
摘要：In view of the existing potential functions' deficiency in the construction of the transition curve based on Metaball technology, a polynomial potential function with a parameter is constructed; then, the application of the potential function in constructing the transition curve is investigated. By properly selecting a class of Bézier curve with parameters, a polynomial potential function with a parameter is ingeniously constructed from the properties at the end points of the curve. Then, the influence of the potential function on the transition curve is investigated, and construction of the optimal transition curve based on the energy optimization method is presented. Examples show the feasibility of the method. The proposed potential function can not only make the transition curve achieve quasi C continuity at the end points but also adjust the shape of the transition curve by modifying the value of the parameter. After the optimal value of the parameter in the potential function is obtained, the smoothest transition curve would be obtained. The proposed potential function alleviates the deficiency of the existing potential functions in constructing the transition curve based on Metaball technology. Furthermore, the construction method of potential function is universal, and the potential functions with different characteristics can be constructed from different curve models.关键词：polynomial potential function;shape adjustment;Metaball technology;transition curve;energy optimization method2512|214|4更新时间：2024-05-07 - 摘要：In view of the proplem of low rate in collision detection of large-scale complex flexible bodies, a new algorithm based on two-phase algorithms is introduced, which are more effective than previous approaches. In the broad phase, we conducted an experiment to construct a 26-DOP bounding volume hierarchy. In the narrow phase, we combined a representative triangle and an orphan set. Subsequently, a new elimination algorithm was introduced. At the filter level, we described the drawback of the non-collinear filter (NCF) and provided a solution. In addition, a new filter named deforming conditional filter (DCF) was proposed and used after DNF and NCF to achieve a high interactive rate. We have implemented our algorithm in some numerical experiments, as described in the second and third parts of the experimental section. for the cloth_ball data set, the use of DNF and NCFI allowed for the number of VF tests to be reduced by 85.90% compared with the use of DNF, whereas the use of DNF, NFCI, and DCF, allowed for a reduction of 87.94%. The proposed approach for general large-scale deformable body collision detection has universality. Particularly in the case of collision processing of triangle flipping, in which DCF and NCF fail, the proposed conditions of the filters can effectively achieve culling and improve the overall efficiency of the algorithm.关键词：large scale deformable objects;collision detection;BVHs;R-Tri;O-Set;filter3231|369|0更新时间：2024-05-07
- 摘要：Physically based fluid animation is an advanced research hotspot in the computer graphics area. However, the realistic and real-time simulation of large-scale fluid simulation is still difficult to handle using current methods. To solve these problems, this study proposes a novel physically based method based on shallow water equation. First, to deal with the instability problems of numerical solution, such as blur and water spot artificial effects, our method proposes to solve shallow water equation using a stable Euler numerical method. Second, we propose a stable model of coupling fluid height field with rigid body and particle system; thus, a two-way fluid-solid coupling and small-scale features can be realistically simulated. Third, to achieve real-time simulation, we design a multiresolution grid algorithm and an interval particle sampling algorithm to accelerate the entire simulation process. Experimental results demonstrate that the proposed method can solve the instability and calculation complex problems and achieve 20 fps in the configuration of 300×300 grid resolution and 22K fluid particles. The proposed method is stable and efficient and suits the realistic simulation of large-scale fluid for real-time applications, such as computer games and 3D vision simulation.关键词：shallow water equation;physically based fluid simulation;Euler method;realistic simulation;real-time2789|302|3更新时间：2024-05-07
- 摘要：Owing to the complexity and randomness of CTA images in the clinic, clinicians are required to manually segment the coronary artery to diagnose a patient suffering from coronary artery diseases. How to segment the coronary artery quickly, accurately, and automatically is important for diagnostic efficiency. According to the features of dual-source CT image, a three-dimensional segmentation algorithm for cardiac coronary artery is investigated. To solve the difficulty of single region or boundary-based active contour model, the segmentation method of active contour model-based vessel prior shape is proposed. First, the improved fuzzy C-means algorithm is used for the initial coarse segmentation of the region of interest (ROI) in cardiac CT images to extract the ROI. The coarse segmentation result is applied to initialize the C-V model level set automatically and estimate the controlling parameters for level set evolution. Next, the 3D heart volume data are used to obtain multiscale gradient vector information and determine the structure of the boundary-based energy function. Then, the vascular structures within the 3D ROI volume are enhanced using the Hessian matrix-based multiscale vascular function. After vascular filtering, the vessel's prior shape information can be obtained to construct the region-based energy function. Finally, the boundary- and region-based energy functions are fused to construct the hybrid energy function, and the variation principle and level set method are used to derive the level set evolution equation for coronary vessel segmentation. The gray scales of vessel images are uneven, and the area of the terminal vessel is smaller than the others; thus, the aforementioned algorithm will be divided into several small-scale subareas of the vessels for the evolution of the contour. The vessel prior information and gradient information are fused to segment accurately the coronary artery from the image with uneven distribution of the gray and contrast agents, better than the conventional vessel segmentation method. Furthermore, good segmentation effect can further be obtained even for a small vessel structure. Experimental results show that the proposed method is only dependent on the initial contour to extract 3D vascular images effectively. Several groups of heart volumes are tested to evaluate its performance in this study. The results show that active contour model-based vessel prior shape can accurately segment the complete coronary structures with a simple human interaction. The method has good accuracy and superiority in terms of the automatic segmentation of a dual-source CT coronary artery image.关键词：coronary artery segmentation;cardiac CT image;active contour model;shape constraint;level set method3165|283|4更新时间：2024-05-07
- 摘要：Several problems exist in polarimetric synthetic aperture radar (SAR) image classification, such as feature selection subjectivity and low utilization efficiency of massive features. Deep belief network (DBN) has a significant advantage in feature learning, which can be used in learning and extracting effective features from massive original features. Based on this observation, a polarimetric SAR image classification method based on DBN is proposed. The proposed method is capable of level-by-level learning and abstracting for the mass original polarimetric features. First, the original polarimetric feature sets are extracted from polarimetric SAR images. Second, 20000 samples are selected, and feature vectors are constructed. Each pixel contains 267 original polarimetric features and class labels. Thus, a pixel is a sample, namely, a feature vector. The feature vector is used as input in the DBN model. Then, the DBN model is built to extract abstract features, namely, effective features. These features are achieved through level-by-level learning. Finally, the logistic regression, a classifier at the top of the DBN model, is applied to classify the entire polarimetric SAR image. Considering AIRSAR data as an example, the overall classification accuracy can reach a high accuracy of 91.06%. The DBN method shows outstanding advantage in feature learning. Simulation experiments show that compared with the traditional Wishart supervised classification algorithm, the DBN algorithm performs much better in classification. Simultaneously, the necessity of the DBN model has been proven by comparing with the logistic regression classification. The logistic regression classification classifies the polarimetric image using the original polarimetric features without any deep learning and extraction. In brief, the effectiveness of the DBN model has been validated through analysis and comparison. In this study, a novel polarimetric SAR image classification method is proposed. Mass polarimetric features of the polarimetric SAR image are utilized for the first time through the DBN. The advantages and applicability of the proposed method are analyzed. Overall, a novel method is proposed for polarimetric SAR image classification, which paves the way for further research and offer beneficial attempts for the utilization of DBN in polarimetric SAR image processing.关键词：polarimetric SAR;deep belief network;image classification;deep learning;remote sensing;feature learning3839|264|6更新时间：2024-05-07
- 摘要：Most hyperspectral remote sensing images suffer from degradation because of the distortion of atmospheric transmission, the limitation of electron devices, and the influence of poor illumination. As a result, the performance of these images in subsequent applications is seriously affected. Thus, the noise in hyperspectral images must be estimated. Given that the noise levels in different bands are often not equivalent in practice, the noise level in each band must be estimated to select an efficient subset of bands. To achieve this end, this paper proposes a hyperspectral image noise estimation algorithm. First, given the high correlation between hyperspectral channels, a low-rank-based model is established specifically for the hyperspectral case. A proper furthermore rule is selected for the noise estimation model to achieve robust performance. Second, the noise in hyperspectral channels is estimated simultaneously using the proposed model. Third, the noise density in each band is calculated as noise level, and the useless bands can be rejected. Experiments are performed on both simulated and real datasets. The proposed method is more robust and can achieve better results than several existing methods because it fully utilizes the correlation and difference between bands. Given that the noise level in bands may be unbalanced, this paper proposes a noise estimation algorithm for hyperspectral images by exploiting the low-rank characteristic of hyperspectral data. By considering noise analysis, this paper proposes a new method to evaluate the quality of hyperspectral images without reference. The proposed algorithm can be applied to highly correlated multichannel images, and the evaluation results are in accordance with expert knowledge and manual interpretation.关键词：hyperspectral image;noise estimation;noisy band detection;low-rank representation3300|238|1更新时间：2024-05-07
- 摘要：The distribution of ships at sea is a key factor for maritime traffic analysis and ship safety management. With the rapid development of earth observation technology, remote sensing is now considered a useful tool to detect ships at sea on a large scale. Particularly owing to the unique technical properties, such as being less vulnerable to cloud and mist and being unaffected by day and night, synthetic aperture radar (SAR) is widely used for ship detection in maritime security management. However, azimuth ambiguities caused by the mechanism of SAR imaging can be easily misclassified as ships on SAR images, leading to a high false alarm rate in ship detection, which has been a difficult problem in ship monitoring with SAR. Considering this issue, the mechanism of azimuth ambiguities on SAR images was initially analyzed in this study. Then, a new method for azimuth ambiguity removal was proposed based on this mechanism. The removal process of azimuth ambiguities includes three steps. First, the consistency of angles is estimated between the real target and its azimuth ambiguities. In this step, the determination method of azimuth angle between real target and its azimuth ambiguities was also discussed. Second, the uniformity of offset distance is determined, and determining the method of the azimuth distance between the real target and its azimuth ambiguities was also discussed. Third, energy decay is analyzed in the azimuth direction, considering that azimuth ambiguities of real ships on SAR images will follow the principles of energy decay. Using these three discriminant criteria, bright targets detected from SAR images can be classified as real ships and azimuth ambiguities. Radarsat-2 images covering the Bohai Sea or the North of the Yellow Sea were selected for a case study; the spatial resolution of these test images captured from March to June 2015 was 30 m. Using the method proposed in this research, azimuth ambiguities of ships were removed step by step and stored in a geodatabase. Real ship targets were further extracted and stored in a geodatabase. These results were compared with the Automatic Identification System data, which can be considered factual data for the case study. Experimental results indicate that all azimuth ambiguities in the study area were detected and removed from real ship targets. After being tested with four Radarsat-2 images, the average accuracy of this azimuth ambiguity removal method based on spaceborne SAR images proposed in this research is more than 95.8%. The results showed that this method can be effectively used to distinguish real ships from its azimuth ambiguities for 30 m spatial resolution SAR images and can improve the accuracy of ship detection on SAR images.关键词：ship detection;azimuth ambiguities;synthetic aperture radar (SAR);automatic identification system(AIS);remote sensing;Radarsat-22599|237|1更新时间：2024-05-07
- 摘要：Currently, downsampling filtration is the main method of generating a spatial image pyramid data, but any objective indicator has not occurred to evaluate the effect of a downsampling filter because calculating the filter's downsampling PSNR requires at least two layers of original data of a spatial pyramid. This study establishes the research technology roadmap of solving this problem:based on two-layer original signals of video image data, an excellent performance downsampling filter was discovered and identified and the subjective visual effect of its generation of the remote-sensing pyramid was verified. Finally, we proposed a downsampling method of filtering along the image texture direction to generate a high-quality spatial pyramid image. Downsampling and upsampling were presented to have been combined to form a pair of resampling filters, as RSFP served as an approximation of the current layer data of pyramids, which can be used to evaluate the downsampling filtering effect. Based on RSFP, a novel pyramid-generating approach TDFA was established:For each 8×8 block, TDFA searches to determine the one-texture orientation of the image from five directions, i.e., DC, horizontal, 135°, vertical, and 45°, and filters along the texture direction implementation of downsampling with a three-order filter, with its effect better than the currently best nearest neighbor interpolation, and the pseudo color, zipper, mosaic, or block effect does not exist. Experimentation and comparisonof effect using a large quantity of image data with several typical downsampling filters, the increasing range of average PSNR by TDFA, was approximately 7.29 dB to 8.44 dB for the Lagrange filter; approximately 6.26 dB to 7.40 dB for the bicubic filter; approximately 5.80 dB to 6.84 dB for AVS's 1/4 interpolation filter; and approximately 4.51 dB to 5.70 dB for the nearest neighbor interpolation method. This study proposed a type of texture-filtering algorithm TDFA. This algorithm can be used to generate the remote sensing pyramid, with its quality better than the level of the existing best generation method, and can also be used to generate a high-quality multilayer video-streaming media data. The proposed resampling filtering pair RSFP can output the high-precision prediction of the current layer of the pyramid used for a scalable video-encoding process.关键词：image pyramid;down-sampling;up-sampling;re-sampling filter pair;texture filtering;evaluation of filter2682|396|1更新时间：2024-05-07
- 摘要：The resolution of unmanned aerial vehicle (UAV) video has a direct effect on target recognition and information acquisition, thereby playing a highly significant role in improving video resolution. Currently, super-resolution reconstruction for UAV is proposed to improve the quality of UAV reconnaissance video for the characteristics of the UAV camera and camera data. The feature matching algorithm based on AGAST-Difference and Fast Retina Keypoint (FREAK) is primarily proposed to match the video object frame and the adjacent frames. Then, the matching region search method is proposed to find the corresponding relationship between the target frame and the aerial image, and aerial photographs are used to make high-frequency compensation of video frame. Finally, solving the optimization of video compensated by the proposed iteration steps utilizes the Projection Onto Convex Sets (POCS) method. Experimental results show that the feature matching algorithm based on AGAST-Difference and FREAK has significant advantages in scale, rotation, and viewpoint. The matching region search method improves the high-frequency compensation continuity of UAV video. POCS iterative optimization improves the reconstruction capability of edge preservation. Compared with the algorithm presented in A Simple & Effective Video Sequence Super-Resolution Algorithm, the algorithm in this study is approximately five times faster, and its image reconstruction is improved by approximately 4 dB. In this study, super-resolution reconstruction for UAV video is presented, and the core of the UAV video super-resolution is analyzed. The feature matching algorithm based on AGAST-Difference and FREAK as well as the matching region search method are proposed to solve problems of image registration and high-frequency compensation. The experimental results show that the consistency and fidelity of the reconstructed image are enhanced, the effect of the image edge detail is especially extremely obvious, and the processing speed is fast.关键词：super-resolution;unmanned aerial vehicle;AGAST-Difference;matching region search method;fast retina keypoint (FREAK);projection onto convex sets (POCS)3647|387|3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