Ship Target Detection in SAR Images Based on Multiple Attention Mechanism and Cross-Scale Feature Fusion

Aiming to address the challenges of inefficient target detection in SAR images caused by complex backgrounds, small ship targets, and significant scale variations, this paper proposes a novel SAR ship target detection model, YOLO-SS, based on YOLOv10n. First, the method incorporates the Bottleneck Transformer (BoTNet) into the Backbone of the baseline network module. This enhancement enables the model to better capture global features and contextual information by establishing global relationships between image regions at different locations, thereby reducing the false and missed detection rates of SAR ship targets in complex environments. Second, an R-BiFPN feature fusion module is introduced in the Neck of the baseline network. This module deeply integrates target features with rich boundary information extracted from the shallow network and highly semantic information captured by the deep network. This fusion mechanism not only improves the model's performance in recognizing complex scenarios, such as small ship targets, fuzzy targets, and dense targets in ports, but also significantly enhances the overall accuracy, robustness, and multiscale perception capability of the model. Furthermore, by incorporating the Normalized Wasserstein Distance (NWD) and combining it with the CIoU loss function, a CIoU-NWD weighted loss function is designed. This reduces the sensitivity of the CIoU loss function to positional offsets of small targets, with only a slight increase in computational and parameter costs, thereby further improving the detection accuracy of small targets. Extensive experiments are conducted on three public datasets—SSDD, HRSID, and SARShip-Dataset—to validate the effectiveness of the proposed method. The experimental results demonstrate that, compared to the baseline model YOLOv10n, the proposed model achieves improvements of 1.2% in mAP50 and 2.1% in mAP50-95 on SSDD, 1.4% in mAP50 and 2.4% in mAP50-95 on HRSID, and 0.9% in mAP50 and 2.0% in mAP50-95 on SAR-Ship-Dataset.