CPC-YOLO: Lightweight Framework for Small Ship Detection in SAR Images

Convolutional Neural Networks (CNNs) have demonstrated potential in Synthetic Aperture Radar (SAR) ship image detection, yet challenges remain: 1) nearshore docks or coastal islands are easily misidentified as small ships; 2) dense ship distributions and multi-scale variations in nearshore areas lead to severe overlapping of predicted bounding boxes, causing missed detections after Non-Maximum Suppression (NMS) due to the loss of valid boxes. To address these issues, we propose CPC-YOLO with three key innovations, namely a P2 detection head that preserves high-resolution features for ships as small as 4× 4 pixels, a C2PSA-CPCA module that integrates local and global context to enhance feature extraction under complex backgrounds, and an Inner-PIoUv2 loss function that combines PIoUv2's adaptive penalty with Inner-IoU's auxiliary box mechanism for precise multi-scale regression. We validate CPC-YOLO on LS-SSDD-v1.0 and SAR-Ship-Dataset through ablation and comparative studies against state-of-the-art methods. The results demonstrate that CPC-YOLO outperforms the baseline by 2.15% and 1.64% in AP on LS-SSDD-v1.0 and SAR-Ship-Dataset, respectively, while achieving faster inference speed, confirming the effectiveness and superiority of our approach.