YOLOv5-SRR: Enhancing YOLOv5 for Effective Underwater Target Detection

Underwater target detection is a crucial aspect of ocean exploration, and advances in related technologies are of significant practical importance. When compared to other target detection tasks, such as those performed on land or in the air, underwater target detection presents a distinct set of challenges. These include poor image quality as a result of light attenuation and water turbidity, as well as the presence of small and densely packed targets that can be difficult to identify. Furthermore, the computational resources available within an underwater vehicle are often limited, further exacerbated the difficulty of detecting and identifying targets in this environment. Despite the excellent performance of existing underwater target detection algorithms with various underwater datasets, these methods often fail to achieve satisfactory results in complex underwater environments where small targets are prevalent. This paper proposes an improved version of the advanced target detection algorithm, YOLOv5, specifically tailored to address this issue. The proposed model incorporates a new SPD-Block module, RepBottleneck structure, and RepBottleneck-ResNets structure, along with the Soft-NMS algorithm. The resulting model, YOLOv5-SRR, demonstrated remarkable effectiveness in underwater target detection, achieving an average accuracy (mAP) of 83.6% on the URPC2020 (Dalian) competition dataset. This result surpassed that of current general target detection models and proved to be more appropriate for use in complex underwater environments.