ABSTRACT Color image restoration is a fundamental task in computer vision and image processing, with extensive real‐world applications. In practice, color images often suffer from degradations caused by sensor noise, optical blur, compression artifacts, and data loss during the acquisition, transmission, or storage. Unlike grayscale images, color images exhibit high correlations among their RGB channels. Directly extending grayscale restoration methods to color images often leads to issues such as color distortion and structural artifacts. To address these challenges, this paper proposes a novel quaternion‐based color image restoration framework. The method integrates low‐rank pseudo‐norm constraints with saturation‐value total variation (SVTV) regularization, effectively enhancing restoration quality in tasks including denoising, deblurring, and inpainting of degraded color images. The proposed algorithm is efficiently solved using the alternating direction method of multipliers (ADMM), and restoration performance is rigorously evaluated through quantitative metrics including peak signal‐to‐noise ratio (PSNR), structural similarity index measure (SSIM), and S‐CIELAB error. Extensive experimental results demonstrate the superior performance of our method compared to existing approaches.

Compared to the horizontal bounding box (HBB), the oriented bounding box (OBB) exhibits substantial advantages in SAR ship detection owing to its outstanding ability to reflect the shape and orientation of ships. However, most OBB detectors are anchor-based, which not only requires the manual setting of numerous hyper-parameters but also increases the difficulty of network training and inference due to the introduction of angle information. Therefore, this paper proposes an anchor-free OBB detector to achieve arbitrary-oriented SAR ship detection. Specifically, the OBB is unified as a 6-distance representation related to each sample point, which is conducive to reducing the negative impact of measurement inconsistency on network training. Moreover, to boost the location accuracy of the ship, a neighborhood-weighted rotation non-maximum suppression method is presented to filter high-quality redundant boxes for corner coordinate weighting. Experiments on RBox-SSDD and HRSID illustrate that the proposed method outperforms some advanced arbitrary-oriented ship detectors.

In real synthetic aperture radar (SAR) scenarios, the scarcity of labeled samples is a common problem, which brings challenges to deep-learning based automatic target recognition (ATR) methods. This work proposes an enhanced prototypical network with multiscale region-aware convolution (MRCEPN) to specially tackle the few-shot SAR ATR problem. We first develop a feature extraction module based on multiscale region-aware convolution, which can adaptively adjust convolutional kernels according to each SAR image’s own feature and make full use of variable spatial information, thus augmenting the capability to extract more discriminative features. Then, an enhanced prototypical network is proposed for label prediction, which can update the class prototype with support and query samples together, thus raising the classification accuracy. Moreover, a hybrid loss is designed to learn a feature space with both inter-class separability and intra-class tightness as much as possible, which helpfully improves the recognition performance. Experiments performed on moving and stationary target acquisition and recognition (MSTAR) dataset demonstrate that the proposed method is competitive with some state-of-the-arts for few-shot SAR ATR tasks.

In realistic synthetic aperture radar (SAR) scenarios, new categories of targets continue to appear over time. It requires the automatic target recognition (ATR) model should continue to accommodate and recognize new categories of targets while maintaining stability on old categories of targets. Therefore, how to strike a glorious balance between plasticity and stability is a matter of wide concern for incremental SAR ATR. In this letter, we propose a new incremental ATR method called dynamic embedding relation distillation network (DERDN), which strives to alleviate the dilemma between plasticity and stability. Specifically, we first develop a dynamic feature embedding model whose parameters can be adjusted adaptively with the data, which is very conducive to continuous generalization to new category targets. To mitigate catastrophic forgetting of model on old categories, we then propose a relation distillation strategy based on exemplar replay to recall knowledge of old categories, thereby improving the stability of the model. Moreover, a hybrid loss that takes into account both label space and embedding space is presented, which ensures inter-class discriminability while enhancing intra-class compactness in the feature space, so as to sustainably accommodate new categories. Evaluation experiments on two benchmark datasets, i.e., moving and stationary target acquisition and recognition (MSTAR) and the synthetic and measured paired labeled experiment (SAMPLE), illustrate that the proposed method is competitive with many state-of-the-arts. The recognition rate of the proposed method is up to 3% higher than that of the best competitor, and its forgetting rate is about 2.5% lower than that of the best competitor.

The existing synthetic aperture radar (SAR) automatic target recognition (ATR) methods have shown impressive results in static scenarios, yet the performance drops sharply as new categories of targets are continuously increased. In response to this problem, this letter proposes a novel ATR method named multi-level adaptive knowledge distillation network (MLAKDN) to achieve incremental SAR target recognition. To be specific, an adaptive weighted distillation strategy is first proposed, which can alleviate the model from forgetting the knowledge of old categories by distilling multi-stage soft label information of old categories at the classification level. Then, a feature distillation method based on gradient maximum criterion is developed to filter and distill discriminative features, so as to further recall more knowledge of old categories at the feature level. Meanwhile, a model rebalancing technique is designed to effectively strike the balance of the model on new categories and old categories. Finally, a weighted incremental classification loss is presented to train the whole model. Experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset and the synthetic and measured paired labeled experiment (SAMPLE) dataset illustrate that the proposed method is superior to some state-of-the-arts for incremental SAR target recognition tasks.

Deep learning-based synthetic aperture radar (SAR) automatic target recognition (ATR) algorithms have achieved outstanding performance under the condition of hundreds or thousands of training samples in recent years. Nevertheless, it is often rare to acquire great quantities of target samples in real SAR application scenarios. This article proposes a novel ATR method called transductive prototypical attention reasoning network (TPARN) to solve the problem of SAR target recognition with only a few training samples. To be specific, a region awareness-based feature extraction model is first developed, which can effectively focus on the target region of interest and suppress the background clutter by embedding direction-aware and position-sensitive information to extract more transferable knowledge. To heighten the discrimination of the sample features, a cross-feature spatial attention module is then proposed following the feature embedding model. Finally, a transductive prototype reasoning method is presented to realize the identity reasoning of the target, which can continuously update each class prototype with training samples and test samples together, thereby improving the classification accuracy. In addition, a marginal adaptive hybrid loss is proposed to obtain a discriminative feature embedding space with intra-class compactness and inter-class divergence, aiming to facilitate subsequent target identity reasoning. Extensive experiments on the moving and stationary target acquisition and recognition (MSTAR) benchmark dataset reveal that the proposed method outperforms some state-of-the-arts under different few-shot SAR ATR tasks.

Data-driven automatic target recognition (ATR) methods have become the mainstream in the synthetic aperture radar (SAR) community at this stage. However, in real SAR application scenarios, the scarcity of training samples is a common problem. Especially in military application scenarios, only a small number of samples of each type of target are usually available. In the case of limited available samples, it is bound to bring challenges to the feature extraction model and classifier inference learning. In this paper, we propose a novel method called adaptive convolutional subspace reasoning network (ACSRNet) to address few-shot SAR target recognition tasks. To be specific, we first present a dynamic-aware convolutional feature embedding network based on the siamese architecture, which can not only learn more transferable knowledge for the few-shot tasks, but also dynamically adjust the convolution kernel according to the input data to extract more discriminative features. To effectively achieve target identity reasoning, we then resort to high-order information of samples, i.e., the idea of adaptive subspace learning, to develop a few-shot subspace classification module, which can online reason the identity of the unknown target through the spanning subspace of training samples. Meanwhile, a dual-loss is designed to train a feature embedding space with intra-class compactness and inter-class divergence, aiming to facilitate subsequent classification. Meta-learning integrating random sampling episode way is introduced into the process of model training to realize few-shot SAR ATR tasks by emulating the human cognitive process. Experimental results on the moving and stationary target acquisition recognition (MSTAR) dataset demonstrate that the proposed method is competitive with some state-of-the-arts for few-shot SAR ATR tasks.

This paper proposes a Bi -similarity prototypical network with capsule-based embedding to solve the problem of few-shot SAR target recognition. The proposed method comprises two procedures, i.e., feature embedding module and Bi-similarity reasoning module. Specifically, we build a feature embed-ding network with capsule operation, which can enable a feature embedding network to extract more informative features by effectively encoding relative spatial relationships between features. To reason the identity of target robustly, we develop a reasoning module based on Bi-similarity metric. Moreover, a mixed loss is proposed to train a discriminative representation space with both intra-class aggregation and inter-class separation. Experimental results on moving and stationary target acquisition and recognition (MSTAR) dataset show that the proposed method is effective and superior to some state-of-art methods in few-shot SAR target recognition tasks.

Sufficient training samples are prerequisite for most existing automatic target recognition (ATR) algorithms to obtain satisfactory recognition performance. Nevertheless, sometimes only a few samples are available in real synthetic aperture radar (SAR) application scenarios. Therefore, this paper proposes a Siamese subspace classification network to address the problem of SAR ATR with insufficient training samples. To be specific, the proposed method first establishes a feature embedding network based on Siamese structure, and leverages contrastive learning to train a low-dimensional representation space with both intra-class compactness and inter-class divergence. A subspace learning-based classifier is then designed to reason the identity of the target. Experimental results on moving and stationary target acquisition and recognition (MSTAR) benchmark data set demonstrate the effectiveness and superiority of the proposed method.

Deep learning-based automatic target recognition (ATR) methods can perform well with sufficient training samples, yet their performance will degrade significantly when the number of training samples available is quite small. Therefore, this paper proposes a multi-task representation learning network to achieve few-shot synthetic aperture radar (SAR) target recognition. On the one hand, the proposed network can perceive the input transformation, identity itself and class discrimination simultaneously. On the other hand, the proposed network can also extract the morphological features of the target and realize the feature refinement by channel attention mechanism. The powerful feature learning ability of the proposed network provides a guarantee for feature extraction under the condition of a few samples. Experiments on moving and stationary target acquisition and recognition (MSTAR) data set validate the effectiveness of the proposed network.

Deep learning has attracted intensive attention in synthetic aperture radar (SAR) automatic target recognition (ATR). Usually, a considerable number of labeled samples are necessary to learn a deep model for obtaining good generalization capability. However, the process of sample labeling is time-consuming and costly. This letter proposes an active self-paced deep learning (ASPDL) approach to SAR ATR. In a nutshell, we first introduce the Bayesian inference into the process of deep model parameter optimization, aiming at learning a robust classification model in the case of a limited number of labeled samples. Next, a cost-effective sample selection strategy is presented to iteratively and actively select the informative samples from a pool of unlabeled samples for labeling. Concretely, high-confidence samples are actively selected through self-paced learning (SPL) way and automatically pseudo-labeled with the current classification model, whereas low-confidence samples are chosen through an active learning strategy and manually labeled. Finally, we update the parameters of the model by minimizing a dual-loss function using a new training set that is constructed by incorporating new labeled samples with original ones. Experiments on the moving and stationary target acquisition and recognition (MSTAR) benchmark data demonstrate that the proposed method can achieve better classification accuracy with relatively few labeled samples compared with some state-of-the-art methods.

In this paper, a class-oriented local structure preserving dictionary learning (CLPDL) algorithm is developed for synthetic aperture radar (SAR) target recognition. Unlike most sparse representation algorithms whose sparse model is predefined via dictionary with atoms being training samples themselves, class-oriented dictionary leaning can derive multiple class-dictionary from training set. To preserve data local structure, a local weighted constraint, i.e., Tikhonov regularization, is introduced into the dictionary learning procedure, which is very helpful for certain challenging scenarios such as configuration recognition and large depression variations. Moreover, to reduce the influence of target aspect sensitivity of SAR image on target recognition, the query sample is represented as a linear combination of class-dictionary to eliminate disturbances. Experimental results on the moving and stationary target acquisition and recognition (MSTAR) database demonstrate the validity of the proposed method.

Attention model and saliency detection has been widely researched in computer vision. This paper contributes to present a novel saliency-based detection algorithm for synthetic aperture radar (SAR) interpretation. Firstly, the input image is segmented into superpixels by simple linear iterative cluster (SLIC) process. Secondly, three features maps are created based on standard deviation, peak value and comparison with background, respectively, and the final saliency map is the fusion of the three feature maps. Lastly, a global threshold is utilized to segment the saliency map and extract target regions. Experimental results with real SAR images demonstrate that the produced saliency map is effective on both highlighting targets and inhibiting background. The proposed method also shows better detection performance and less time-consuming than the conventional CFAR.

In this paper, a novel method named discriminant neighborhood preserving projections using L1-norm maximization (DNPP-L1) is developed for Synthetic Aperture Radar (SAR) target recognition. The proposed method can preserve the local geometry information from raw high-dimension data and utilize useful class discriminant information to improve the performance of target recognition effectively. The proposed DNPP-L1 is based on L1 norm distance metric, which is very robust for SAR images target with noise. Experimental results on MSTAR database demonstrate the effectiveness of the proposed method.

In this paper, a new subspace learning algorithm, called enhanced kernel uncorrelated discriminant nearest feature line analysis i¼ˆEKUDNFLAi¼‰, is presented. The aim of EKUDNFLA is to seek a feature subspace in which the within-class feature line (FL) distances are minimized and the between-class FL distances are maximized simultaneously. At the same time, an uncorrelated constraint is imposed to get statistically uncorrelated features, which contain minimum redundancy and ensure independence, and thus it is highly desirable in many practical applications. Optimizing an objective function in a kernel feature space, nonlinear features are extracted. In addition, a weighting coefficient is introduced to adjust the proportion between within-class and between-class information to get an optimal effect. Experimental results on radar target recognition with measured data demonstrate the effectiveness of the proposed method.

In this letter, a new manifold learning algorithm, called uncorrelated discriminant locality preserving projections (UDLPP), is proposed. The aim of UDLPP is to preserve the within-class geometric structure, while maximizing the between-class distance. By introducing a simple uncorrelated constraint into the objective function, we show that the extracted features via UDLPP are statistically uncorrelated, which is desirable for many pattern analysis applications. Moreover, UDLPP can be performed in reproducing kernel Hilbert space, which gives rise to kernel UDLPP. Experimental results on both face recognition and radar target recognition demonstrate the effectiveness of the proposed algorithm.

In this letter, we present a new formulation for uncorrelated discriminant analysis (UDA) in some high-dimensional feature space and then propose an efficient UDA algorithm using kernel technique. Unlike some existing UDA algorithms, which solve uncorrelated discriminant vectors one at a time, the proposed algorithm is able to extract all the uncorrelated discriminant vectors simultaneously in the feature space and does not suffer the small sample size problem. Experimental results show that the proposed method is very competitive in comparison with some existing discriminant analysis algorithms, in terms of recognition rate and robustness with respect to kernel parameters.