Most deep-learning-based target detection methods have high computational complexity and memory consumption, and they are difficult to deploy on edge devices with limited computing resources and memory. To tackle this problem, this article proposes to learn a lightweight detector named Light-YOLOv4, and it is obtained from YOLOv4 through model compression. To this end, first, we perform sparsity training by applying L1 regularization to the channel scaling factors, so the less important channels and layers can be found. Second, channel pruning and layer pruning are enforced on the network to prune the less important parts, which could significantly reduce network's width and depth. Third, the pruned model is retrained with a knowledge distillation method to improve the detection accuracy. Fourth, the model is quantized from FP32 to FP16, and it could further accelerate the model with almost no loss of detection accuracy. Finally, to evaluate Light-YOLOv4's performance on edge devices, Light-YOLOv4 is deployed on NVIDIA Jetson TX2. Experiments on the SAR ship detection dataset (SSDD) demonstrate that the model size, parameter size, and FLOPs of Light-YOLOv4 have been reduced by 98.63%, 98.66%, and 91.30% compared with YOLOv4, and the detection speed has been increased to 4.2×. While the detection accuracy of Light-YOLOv4 is only slightly reduced, for example, the mAP has only reduced by 0.013. Besides, experiments on the Gaofen Airplane dataset also prove the feasibility of Light-YOLOv4. Moreover, compared with other state-of-the-art methods, such as SSD and FPN, Light-YOLOv4 is more suitable for edge devices.

Synthetic aperture radar (SAR) target recognition faces the challenge that there are very little labeled data. Although few-shot learning methods are developed to extract more information from a small amount of labeled data to avoid overfitting problems, recent few-shot or limited-data SAR target recognition algorithms overlook the unique SAR imaging mechanism. Domain knowledge-powered two-stream deep network (DKTS-N) is proposed in this study, which incorporates SAR domain knowledge related to the azimuth angle, the amplitude, and the phase data of vehicles, making it a pioneering work in few-shot SAR vehicle recognition. The two-stream deep network, extracting the features of the entire image and image patches, is proposed for more effective use of the SAR domain knowledge. To measure the structural information distance between the global and local features of vehicles, the deep Earth mover’s distance is improved to cope with the features from a two-stream deep network. Considering the sensitivity of the azimuth angle in SAR vehicle recognition, the nearest neighbor classifier replaces the structured fully connected layer for $K$ -shot classification. All experiments are conducted under the configuration that the SARSIM and the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset work as a source and target task, respectively. Our proposed DKTS-N achieved 49.26% and 96.15% under ten-way one-shot and ten-way 25-shot, whose labeled samples are randomly selected from the training set. In standard operating condition (SOC) as well as three extended operating conditions (EOCs), DKTS-N demonstrated overwhelming advantages in accuracy and time consumption compared with other few-shot learning methods in $K$ -shot recognition tasks.

Deep learning has been widely used in automatic target recognition (ATR) for synthetic aperture radar (SAR) recently. However, most of the studies are based on the network structure in optical images and lack full consideration of the inherent characteristics of SAR targets, which limits the improvement of recognition accuracy and makes poor generalization ability. In addition, due to the black-box characteristics, it is difficult to effectively interpret SAR ATR results. To conquer these problems, we propose an electromagnetic scattering feature (ESF) module embedded network based on attributed scattering center (ASC) model to incorporate the SAR targets’ characteristics into the deep learning framework. First, a novel convolutional neural network (CNN)-based algorithm for extracting ASC parameters is proposed, which makes the network focus on target features under the guidance of physical model. Then, the ESF module is designed based on a well-trained ASC parameters extractor to inject the learned target features into the classification network for more robust and interpretable results. Besides, two structures are proposed combined with the ESF module for single-view and multiview SAR target classification, which further illustrates the portability of the module. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset show the validity of the proposed CNN-based ASC extractor and the ESF module embedded classification network. Compared with ordinary networks, our method can achieve higher classification accuracy under complex conditions, which reflects the better generalization performance of the algorithm. Furthermore, through visualization analysis of the classification results, we show the interpretability of the network combined with the electromagnetic scattering characteristics.

Automatic target recognition of synthetic aperture radar (SAR) images has been a vital issue in recent studies. The recognition methods can be divided into two main types: traditional machine learning methods and deep-learning-based methods. For most traditional machine learning methods, target features are extracted based on electromagnetic scattering characteristics which are interpretable and stable. However, the extraction process of effective recognition features is often complex and the computational efficiency is low. Compared with the traditional methods, the deep learning methods can directly learn the high-dimensional features of the target to obtain higher target recognition accuracy. However, these algorithms have poor generalization performance and are difficult to explain. In order to comprehensively consider the advantages of the two kinds of methods, this article proposes a novel method for SAR target classification based on integration parts model and deep learning algorithm. First, part convolution and modified bidirectional convolutional-recurrent network are used to extract local feature of target through parts model which is calculated based on attribute scattering centers. Then, modified all-convolutional networks are used to extract the global feature of the target. The final classification result is achieved through decision fusion of local and global features. Experimental results on the moving and stationary target acquisition and recognition show the superiority of the proposed method, especially under complex conditions. Besides, a brief analysis of target key parts with part occlusion method is given, which is helpful to the interpretability of the deep learning network.

Machine learning methods for synthetic aperture radar (SAR) image automatic target recognition (ATR) can be divided into two main types: traditional methods and deep learning methods. The deep learning methods can learn the high-dimensional features of the target directly, and usually obtain high target recognition accuracy. However, they lack full consideration of SAR targets’ inherent characteristics resulting in poor generalization and interpretation ability. Compared with the deep learning methods, traditional methods can get more interpretable and stable results with model-based features. In order to take full advantage of these two kinds of methods, we propose target part attention network based on the attributed scattering center (ASC) model to integrate the electromagnetic characteristics with the deep learning framework. Firstly, considering the importance of scattering structure for SAR ATR, we design a target part model based on ASC model. Then, a novel part attention module based on Scaled Dot-Product Attention mechanism is proposed, which directly associates the features of target parts with the classification results. Finally, we give the derivation method of the importance of each part, which is of great significance for practical application and the interpretation of SAR ATR. Experiments on the MSTAR data set demonstrate the effectiveness of the proposed part attention network. Compared with existing studies, it can achieve higher and more robust classification accuracy under different complex conditions. Furthermore, combined with the importance of parts, we constructed two effective interpretable analysis methods for deep learning network classification results.

Most of the existing synthetic aperture radar (SAR) automatic target recognition (ATR) methods aim at the closed set situation, in which the classes of targets in the test set have appeared in the training set. However, in practice, the classifier is likely to encounter the targets from unseen categories and classify them incorrectly, which brings a huge challenge to current SAR ATR techniques. To overcome this problem, this letter proposes an open set recognition (OSR) method based on multitask learning, and the method is developed from generative adversarial network (GAN). Essentially, this method decomposes OSR into two tasks: classification and abnormal detection. The classification task is the same as that in the closed set situation, while the abnormal detection task is used to determine whether the targets belongs to the unseen categories. Correspondingly, the network structure of GAN is modified and the other full-connection network branch is added to the end of the discriminator, so it has the ability to accomplish the above two tasks. Finally, according to the results of two tasks, the OSR for SAR targets can be realized. The experimental results on moving and stationary target acquisition (MSTAR) dataset demonstrate that the proposed method has the better recall, precision, $F1$ , and accuracy than other OSR methods.