Jiang Hongjun

Dalian Ocean University

Abstract：

The intelligent perception of underwater biological targets is a core technology for the automation of marine ranching.However, complex underwater environments, characterized by light attenuation, water turbidity, and the occlusion of aquatic organisms, pose significant challenges to visual object detection.Traditional detection algorithms often suffer from low accuracy and poor real-time performance when dealing with small-sized and overlapped shellfish targets. To address these issues, this paper proposes a high-precision underwater target detection method, named YOLOv8-UW, based on an improved YOLOv8 architecture. Firstly, a Coordinate Attention (CA) mechanism is integrated into the backbone network to enhance the model's sensitivity to critical features while suppressing background noise caused by suspended particles. Secondly, the Content-Aware ReAssembly of Features (CARAFE) operator is introduced to replace the standard nearest-neighbor upsampling in the neck network, thereby improving the feature reconstruction capability for small targets like scallops and sea urchins. Finally, the EIoU (Efficient Intersection over Union) loss function is employed to optimize the bounding box regression, accelerating convergence and improving localization accuracy. Experimental results on the URPC2021 dataset demonstrate that the proposed method achieves a mean Average Precision (mAP@0.5) of 84.6%, which is 3.2% higher than the baseline YOLOv8n model and outperforms other state-of-the-art detectors such as YOLOv7 and SSD. The proposed method effectively balances accuracy and speed, providing a viable technical solution for autonomous underwater harvesting robots.

Key Words：

underwater object detection; deep learning; YOLOv8; attention mechanism; marine ranching; shellfish classification