FFC-YOLO: a SAR ship detection network with enhanced feature representation and dynamic cross-scale fusion

Abstract Ship detection in synthetic aperture radar (SAR) images is crucial for maritime security and operations. However, existing methods face significant challenges when dealing with complex backgrounds and small ships. Ship targets are often misclassified as background clutter, and small ships are particularly susceptible to noise interference. To address these challenges, this paper proposes FFC-YOLO, an enhanced ship detection network based on YOLOv8. FFC-YOLO incorporates several novel modules to improve detection accuracy. Firstly, the Feature Attention Aggregation Module (FAAM) employs multi-scale fusion and Context Anchor Attention (CAA) to enhance detection in complex backgrounds by extracting features from different scales and integrating contextual information. Secondly, the Feature Augmentation Unit (FAU) enhances small ship detection accuracy by employing an improved deformable convolution (A-Defconv) for shape adaptation and an adaptive weighted fusion mechanism to refine feature extraction. Additionally, the Cross-Scale Dynamic Fusion Module (CSDF) dynamically integrates features from different scales, addressing information conflicts that arise when handling multi-scale ship targets, further enhancing detection accuracy. Finally, a novel bounding box loss function, CNW-loss, optimizes the overall network performance. Evaluated on the SSDD and HRSID datasets, FFC-YOLO achieves a significant performance advantage over existing methods, with mAP0.5 scores of 98.5% and 89.7%, respectively. The proposed FFC-YOLO network effectively enhances ship detection accuracy in SAR images, making it highly applicable to various maritime safety and operations, and providing valuable insights for a broader range of measurement applications.