Underwater tiny object detection network based on multi-scale attention and adaptive feature fusion

Abstract

Underwater object detection has long been severely affected by factors such as illumination conditions, suspended particles, and turbidity, which lead to weakened object textures, color distortion, and obscure morphological features. Existing deep learning detection frameworks based on visual features still exhibit unstable performance in these scenarios. To improve model adaptability, this paper proposes Bidirectional weighted Concat with Efficient multi-scale Attention You Only Look Once (BCEA-YOLO), a detection framework designed to mitigate underwater feature degradation through coordinated multi-scale processing. It employs an efficient multi-scale attention mechanism to suppress background noise while utilizing a bidirectional weighted concatenator to refine the fusion of visual features. To better capture tiny and occluded objects, a four-scale detection head with a P2 detection branch is implemented to preserve high-resolution spatial cues. Experimental results on two public benchmark datasets (URPC2018 and URPC2020) demonstrate that the proposed method outperforms competing methods, achieving mean average precision values of 78.5% and 85.5%, respectively, at an intersection over union threshold of 0.5. These results indicate the effectiveness of the proposed method for underwater object detection. Furthermore, edge deployment tests on the NVIDIA Jetson platform validate its real-time inference efficiency, providing a practical visual perception solution for autonomous underwater vehicles.