There are inconsistent tasks and insufficient training in the SAR ship detection model, which severely limit the detection performance of the model. Therefore, we propose a twin branch network and design two loss functions: regression reverse convergence loss and classification mutual learning loss. The twin branch network is a simple but effective method containing two components: twin regression network and twin classification network. Aiming at the inconsistencies between training and testing in regression branches, we propose a regression reverse convergence loss (RRC Loss) based on twin regression networks. This loss can make multiple training samples in the twin regression branch converge to the label from the opposite direction. In this way, the test distribution can be closer to the training distribution after processing. For inadequate training in classification branch, Inspired by knowledge distillation, we construct self-knowledge distillation using a twin classification network. Meanwhile, our proposed classification mutual learning loss (CML Loss) enables the twin classification network not only to conduct supervised learning based on the label but also to learn from each other. Experiments on SSDD and HRSID datasets prove that, compared with the original method, the proposed method can improve the AP by 2.7–4.9% based on different backbone networks, and the detection performance is better than other advanced algorithms.

With the wide application of synthetic aperture radar in maritime surveillance, a ship detection method has been rapidly developed. However, there is still a key problem common in most methods, i.e., how to select positive and negative samples. The mainstream MaxIoUAssign has inherent problems, such as a fixed threshold and rough classification, resulting in the low quality of the positive samples. To solve these problems, we propose a new sample selection method called adaptively center-shape sensitive sample selection. The proposed method introduces shape similarity between proposal boxes and ground truth as one of the evaluation criteria and collaborates with intersection of union (IoU) to measure the quality of the proposal boxes. Meanwhile, the center distance between proposal boxes and ground truth is used to control the influence degree of IoU and shape similarity. In this way, the quality score of the proposal boxes can be determined through IoU, shape similarity, and center position, making sample selection more comprehensive. Additionally, to avoid a fixed threshold, the standard deviation of the quality score is used as a variable to form the adaptive threshold. Finally, we conducted extensive experiments on the benchmark SAR ship detection dataset (SSDD) and high-resolution SAR images datasets (HRSID) datasets. The experimental results demonstrated the superiority of our method.

Deep learning has been widely applied to ship detection in Synthetic Aperture Radar (SAR) images. Unlike optical images, the current object detection methods have the problem of weak feature representation due to the low object resolution in SAR images. In addition, disturbed by chaotic noise, the features of classification and location are prone to significant differences, resulting in classification and location task misalignment. Therefore, this paper proposes a novel SAR ship target detection algorithm based on Cross-Attention Mechanism (CAM), which can establish the information interaction between the classification and localization task and strengthen the correlation between features through attention. In addition, to suppress the noise in multi-scale feature fusion, we designed an Attention-based Feature Fusion Module (AFFM), which uses the attention information between channels to perform the re-weighting operation. This operation can enhance useful feature information and suppress noise information. Experimental results show that on a benchmark SAR Ship Detection Dataset (SSDD), the Fully Convolutional One-Stage Object Detector (FCOS) with ResNet-50 backbone network was optimized to improve AP by 6.5% and computational cost by 0.51%. RetinaNet with ResNet-50 backbone network was optimized to improve AP by 1.8% and computational cost by 0.51%.