UWA-YOLO: Design and Evaluation Protocol for a Lightweight Attention-Guided YOLO Detector for Degraded Underwater Imagery

In this study, an underwater object detection framework named as UWA-YOLO is proposed. The UWA-YOLO has four main parts: (1) Cross-Scale Adaptive Feature Fusion Network (CAFFN), which adaptively weights and exchanges information between shallow and deep feature maps to alleviate the problem of information loss of small objects; (2) Dynamic Attention Detection Head (DADH), which improves localisation and classification confidence of overlapping and dense distribution objects; (3) Underwater Adaptive Image Enhancement Module (UAIEM), which performs learned colour correction and contrast restoration while maintaining object boundaries; and (4) Lightweight Multi-Scale Attention Backbone (LMAB), which combines efficient convolutional blocks with local and global attention to extract multi-scale context. Evaluation is conducted on the established URPC underwater benchmark, with comparisons against Faster R-CNN, SSD, YOLO models, and representative DETR-based and SwinTransformer-based underwater detectors, using precision, recall, F1-score, mAP@0.5, and mAP@0.5:0.95.