YOLO-UOD: Location-Edge Co-Perception Network With Multiscale Features Fusion for Underwater Object Detection

Underwater object detection (UOD) plays a vital role in marine resource exploration and ecological conservation. However, underwater environments pose significant challenges, including non-uniform light attenuation, color distortion between marine organisms and backgrounds, and severe image degradation. These factors lead to blurred boundaries, texture loss, and difficulties in distinguishing objects from their surrounding backgrounds, particularly for small, occluded, and camouflaged objects. To address these issues, we propose a Location-Edge Co-Perception Network with multi-scale features fusion (YOLO-UOD) for underwater object detection. First, we design a Wavelet-Cross Stage Partial (CSP) Synergistic Feature Enhancement Module (WCS-M) that guides CSP operations with discrete wavelet transform (DWT) to enhance feature representation, effectively distinguishing foreground from background. Second, we propose a Location Perception Module (LPM) and an Edge Perception Module (EPM) to enhance semantic preservation and edge-texture awareness using deformable convolutions and adaptive attention mechanisms. Furthermore, a Partial Bidirectional Spatial Attention (PBSA) Module is proposed to enhance target localization and recognition performance while effectively suppressing high-frequency noise. Extensive experiments on benchmark datasets (DUO, RUOD, and UTDAC2020) demonstrate that the proposed method effectively alleviates issues such as blurred boundaries and texture loss in UOD. The key modules also significantly improve performance compared to existing YOLO-series methods, showcasing strong generalization ability. The model code is available at https://github.com/liuyuanjie123-tech/YOLO-UOD.