Abstract
With the rapid advancement of intelligent aquaculture, precise individual identification of underwater fish has become a crucial method for achieving smart farming. By accurately recognizing and tracking individuals within the same species, researchers can enable individual-level identification and tracking, significantly enhancing the efficiency of research and management. To address the challenges of complex underwater environments and subtle differences among similar individuals that affect recognition accuracy, this paper proposes a fish individual identification method based on global detail enhancement and keypoint region fusion. This method simultaneously learns global refined features and keypoint region features, dynamically capturing effective keypoint features while mitigating errors caused by noise through weighted fusion. The network first employs a global detail enhancement module to extract global features, such as overall morphology and texture information, followed by the extraction of fine-grained features from keypoint regions. Through the weighted fusion, the network further emphasizes critical areas, thereby enhancing robustness and adaptability in complex underwater scenarios. This design effectively integrates global refined features and local keypoint features, providing comprehensive support for accurate fish individual identification. Experimental results show that the proposed method achieves mAP and Rank-1 scores of 89.7% and 95.3%, respectively, and demonstrates strong generalization capabilities in other fish identification tasks.