Optimized YOLOv8s framework with deformable convolution for underwater object detection

The booming aquatic economy drives demand for underwater object detection, which faces key challenges: small/occluded targets, variable object morphologies, and turbidity-induced low image quality. To address these, this paper proposes O-YOLOv8s-DC-an optimized YOLOv8s framework for deep learning-based underwater object detection. It integrates four core enhancements: a deformable convolution feature module (C2f_DC, adapting to shape/size variations), a depth-weighted bidirectional feature pyramid (DeepBiFPN, boosting small-target detection), content-aware feature reorganization (CARAFE, reducing occluded-target detection errors), and efficient multi-scale attention (EMA, suppressing redundant features). Ablation studies confirm individual module effectiveness. Experiments on the LFIW and OI datasets show O-YOLOv8s-DC outperforms mainstream models (e.g., SSD, original YOLOv8s, DETR), with AP@[0.50:0.05:0.95] (a comprehensive detection metric) significantly higher than YOLOv8s and occluded-target performance effectively enhanced at strict IoU thresholds (e.g., AP@0.75). It also optimizes small-target recognition accuracy, enabling reliable detection in complex underwater environments and providing technical support for aquatic ecological protection and sustainable underwater operations. Source code: https://github.com/WangXin81/O-YOLOv8s-DC .