Small Ship Detection in SAR Images Based on Asymmetric Feature Learning and Shallow Context Embedding

Ship detection using Synthetic Aperture Radar (SAR) imagery is a critical technology with extensive applications in both civilian and defense domains. While deep learning has considerably advanced SAR ship detection, the identification of small ship targets remains challenging due to their limited size, weak feature representation, and high sensitivity to Intersection over Union (IoU) metrics. These factors hinder detection networks from effectively learning discriminative representations, resulting in performance degradation. To address this issue, we propose a small ship detection network based on the YOLO architecture, integrating asymmetric feature learning and shallow contextual embedding. First, an Asymmetric Feature Extraction (AFE) module with pinwheel-shaped convolutions is introduced to flexibly capture features of small targets in the backbone network. Second, a Shallow Auxiliary Fusion (SAF) module is designed to embed and efficiently fuse essential contextual information from shallow feature maps. Third, an Asymmetric Shuffle Detection Head (ASDet) is proposed to enhance feature interpretation through asymmetric channel blending. Finally, Normalized Wasserstein Distance (NWD) loss is adopted to alleviate the high IoU sensitivity of small targets to bounding box deviations. Extensive experiments on the SSDD and HRSID datasets demonstrate the effectiveness of the proposed method, achieving improvements of 1.7% in small-ship detection accuracy on SSDD and 1.6% on HRSID, outperforming baseline models.