CDDS-YOLO: a sonar image object detection method based on hierarchical dynamic perception and spatial enhancement

Abstract

Side-scan sonar image interpretation plays a critical role in seabed exploration and underwater target inspection; however, complex backgrounds, blurred boundaries, and weak texture features pose significant challenges to robust object detection. To address these issues, this paper proposes a novel detection framework termed CDDS-YOLO, which incorporates hierarchical dynamic perception and spatial enhancement mechanisms for improved sonar image analysis. Built upon the YOLOv11n baseline, the proposed model integrates a sonar-oriented collaborative dynamic-deformable network backbone, a spatially enhanced background decoupling (SEBD) head, and an acoustically motivated composite loss function combining WIoU, focal loss, and distribution focal loss. The backbone improves feature extraction under slant-range distortion, irregular acoustic shadows, and echo-intensity variations; the SEBD head introduces spatial enhancement attention mechanism into the classification branch to suppress reverberation- and speckle-induced background activations; and the composite loss enhances hard-sample learning and blurred-boundary localization under low-contrast conditions. Comprehensive experiments conducted on the S-RAPP dataset demonstrate that CDDS-YOLO achieves up to 90.3% mAP@50 in the comparative experiment and maintains a stable performance of 89.8 ± 0.7% mAP@50 across repeated ablation runs, outperforming several state-of-the-art detectors while maintaining a lightweight architecture. Ablation studies further verify the effectiveness of each proposed module. The results indicate that CDDS-YOLO exhibits strong robustness to boundary ambiguity and complex seabed textures, making it well suited for practical sonar-based detection tasks. Overall, the proposed CDDS-YOLO achieves accurate and robust underwater target detection, demonstrating its effectiveness for sonar image analysis.