In the field of synthetic aperture radar (SAR) vehicle detection, detectors based on deep learning have gained extensive application in recent years. However, most methods can only generate horizontal bounding boxes (HBBs). Due to the lack of angle information, the presence of excessive background noise in HBB leads to an inaccurate description of vehicles. Although research has been conducted on oriented SAR target detection, several issues persist, including the inadequate utilization of vehicle characteristics, inaccurate feature extraction for vehicles exhibiting size and rotation variations, and imprecise positioning of oriented bounding boxes (OBBs). To address these issues, we propose a novel-oriented SAR vehicle detection network. First, leveraging the shape property of vehicles, we devise a constrained rectangular-invariant rotatable convolution (Cons-RIRConv) for the backbone. Under the constraints of a carefully designed size- and rotation-invariant regularization term, Cons-RIRConv determines convolution sampling positions that are adaptive to the size and rotation variations of vehicles. This approach is conducive to Cons-RIRConv in extracting size- and rotation-invariant features of vehicles. Furthermore, as the sampling positions of Cons-RIRConv align closely with vehicles, we integrate them into the anchor generation process within the detection head, thereby developing an adaptive anchor generation mechanism (AAGM). Guided by the sampling positions of Cons-RIRConv, the anchor boxes generated by AAGM exhibit a high degree of conformity with the vehicles and low redundancy, ultimately improving the localization accuracy of OBB and reducing computation costs. Experiments on three authoritative measured SAR vehicle detection datasets show the effectiveness of our method.

Deep-learning-based methods have recently achieved significant advancements in synthetic aperture radar automatic target recognition (SAR ATR). However, these methods typically rely heavily on extensive annotations, which are difficult to obtain for SAR images. Semi-supervised learning offers a solution to improve model performance with limited labeled data by leveraging unlabeled data. The mainstream semi-supervised learning methods for SAR ATR typically select high-confidence unlabeled images to assign pseudo-labels for their inclusion in the model training process. However, the large number of low-confidence unlabeled images are not efficiently utilized. To address this issue, a semi-supervised SAR target recognition method based on contrastive learning and complementary label (CoL) learning is proposed. First, CoL learning assigns CoLto low-confidence unlabeled images based on their minimum prediction probabilities. Subsequently, a threshold is set to filter out unreliable CoL, thereby mitigating the adverse effects of erroneous CoL. This approach ensures the effective and comprehensive utilization of low-confidence unlabeled images. Additionally, we propose a contrastive loss that incorporates CoL. Compared to traditional contrastive losses, our proposed contrastive loss constructs a richer set of negative sample pairs by leveraging the characteristics of CoL more effectively. Consequently, this approach improves the utilization of low-confidence images and further improves recognition performance. In contrast to the current state-of-the-art semi-supervised recognition methods, experiments on the MSTAR dataset demonstrate the better recognition performance of our proposed method with limited labeled images.

This article delves into the topic of unsupervised domain adaptation (UDA) by transferring knowledge from rich labeled optical domain to unlabeled synthetic aperture radar (SAR) domain, tackling the issues faced by deep-learning-based SAR ship classification methods that rely on abundant labeled SAR images. Typical UDA methods usually extract domain-invariant representations (DIRs) between two domains. However, due to the prominent differences in imaging mechanisms between optical and SAR images, the discriminative characteristics of same classes across domains may vary. Feature representation guided by labeled optical images therefore suffers from a particularly serious source-bias problem, making DIR difficult to be extracted. Moreover, capturing the category structure of the target domain is crucial for classification tasks. To solve the above challenges, this article proposes a UDA framework for SAR ship classification via progressive feature alignment between optical and unlabeled SAR domains, gradually aligning two domains across domain and class levels. At the domain level, to reduce the transfer difficulty stemming from the prominent differences between SAR and optical images, feature calibrated domain alignment (FCDA) is presented to achieve accurate DIR extraction. FCDA combines the reconstruction and the consistency constraints of different perturbed versions of the same image to calibrate the optical-bias representation into the features of unbiased toward a specific domain. At the class level, we proposed feature enhanced class alignment (FECA) to capture the fine-grained category structure of the SAR domain. FECA incorporates pseudo-label-based cross-domain contrastive learning (CDC) for intraclass compactness as well as interclass separation among cross-domain categories, along with a consistency learning approach to enhance the class structure of SAR domain. The experimental results indicate that our method achieves exceptional performance in unsupervised classification of SAR ships.

In recent years, target detection methods based on deep learning have achieved extensive development in synthetic aperture radar (SAR) ship detection. However, training such detectors requires target-level annotations of SAR images that are hard to be obtained in practice. To reduce the dependence of network training on expensive target-level annotations, we propose a novel semisupervised SAR ship detection network via scene characteristic learning. The proposed network focuses on utilizing the scene-level annotations of SAR images to improve the detection performance in the case of limited target-level annotations. Compared with the traditional fully supervised SAR ship detection network, the proposed network constructs a scene characteristic learning branch parallel with the detection branch. In the scene characteristic learning branch, a scene classification loss and a scene aggregation loss are designed to utilize the scene-level annotations. Under the constraint of these two losses, the feature extraction network can fully learn the scene characteristics of SAR images, thus enhancing its feature representation ability for ship targets and clutter. In addition, we propose a hierarchical test process from scene to target. After recognizing the scene types of input SAR images, we design different detection strategies for SAR images recognized as different scenes. The proposed test process can significantly reduce the inland and inshore false alarms, thus leading to higher detection performance. The experiments based on two measured SAR ship detection datasets demonstrate the effectiveness of the proposed method.

In this letter, a target detector based on gradient harmonized mechanism (GHM) and attention mechanism is proposed to realize synthetic aperture radar (SAR) target detection in complex scenes. Considering the imbalance of positive and negative examples in SAR target detection, we use RefineDet as our backbone network. RefineDet can mitigate this imbalance problem by introducing the idea of two-step classification and regression into the one-stage detector. However, RefineDet only selects a part of examples for training and does not make full use of the information of all examples. Therefore, we apply GHM to the classification loss function of RefineDet, so that the network can make full use of all examples and increase the weights of hard examples adaptively in the loss function to reduce the false alarms and the missing alarms. In addition, to achieve a better detection performance in SAR images with complex scenes, a multiscale feature attention module (MFAM) is embedded into the network. By applying channel and spatial attention mechanisms to the multiscale feature maps, the MFAM can highlight the significant information and suppress the interference caused by clutter. The extensive experimental results based on the measured SAR dataset verify the effectiveness of the proposed method.

In recent years, Synthetic Aperture Radar (SAR) ship detection methods based on convolutional neural networks have attracted wide attention in remote sensing fields. However, these methods require a large number of labeled SAR images to train the network, where labeling for SAR images is more expensive and time-consuming than that for optical images. To address the problem of lacking labeled SAR images, in this paper, we proposed an unsupervised domain adaptation framework based on progressive transfer for SAR ship detection by transferring knowledge from the optical domain to the SAR domain. Due to the prominent difference between the optical and SAR images, our approach progressively transfers knowledge from three levels: pixel level, feature level and prediction level. At the pixel level, considering the difference in imaging mechanism, we propose a special data augmentation method for ship targets and build the generator with skip-connection based on generative adversarial networks (GANs) to generate transition domain, which can reduce the appearance discrepancy between the optical and SAR images. At the feature level, the detector is trained to learn the domain-invariant features with adversarial alignment. At the prediction level, we further use the predicted pseudo-labels obtained by the feature-aligned detector to learn more discriminative features of the SAR images directly and propose the robust self-training (RST) method to reduce the influence of noisy pseudo-labels on detector training. Specially, RST is formulated as a loss minimization problem for object detection. The experimental results based on the domain adaptation from optical dataset to SAR dataset demonstrate that our approach achieves superior SAR ship detection performance with unlabeled SAR images.

The Single Shot multi-box Detector (SSD) has been successfully applied in synthetic aperture radar (SAR) target detection. Besides, the saliency information in the saliency map has the ability that strengthening the target of interest and suppressing the clutter. It will help to improve the capability of scene understanding. According to the above, a novel SAR target detection network based on saliency-combined SSD is proposed. The proposed method includes two backbone sub-networks, one taking the SAR images as input for extracting the features, and the saliency map obtained from traditional saliency method is used as the input of the other sub-network to obtain the refined saliency information. Through the fusion module is used in multiple scales to integrate the saliency information and the network feature. Finally, we can get the detection results by the convolutional predictors on the multi-scale integrated feature maps. In addition, we apply the dense connection structure in the two sub-networks to utilize context information. The experimental results based on the miniSAR real data show that the proposed method can achieve a good detection performance.