Overcoming body occlusion challenges: a novel underwater fish individual recognition based on local and global feature fusion

• Published in: Aquaculture international Vol. 33; no. 6; p. 512
• Main Authors: Liang, Xuelan, Wu, Junfeng, Xie, Yongzhi, Chen, Miao, Cheng, Shaojiang, Sha, Yida, Wu, Peihua, Zheng, Saibei, Yu, Hong
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.11.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Aquaculture; Aquatic plants; Biomedical and Life Sciences; Coral reefs; Datasets; Deep learning; Deformation; Efficiency; Fish; Fish culture; Fisheries management; Fishery management; Freshwater & Marine Ecology; Identification; Life Sciences; Localization; Machine learning; Methods; Population density; Sustainable fisheries; Underwater; Zoology; Occlusion; Deep learning; Deep orthogonal local and global feature fusion models; Individual fish identification; Complex underwater environment
• ISSN: 0967-6120; 1573-143X
• DOI: 10.1007/s10499-025-02160-z

With the advancement of intelligent aquaculture, accurate fish individual recognition has become critical for monitoring growth, behavior, and health. However, the complex underwater environment, including occlusion, deformation, and dense populations, presents significant challenges to robust identification. In this paper, we propose a fish recognition method based on orthogonal fusion of local and global features to address these issues. By integrating a deep feature fusion network with the YOLOv9 detection framework, our approach enables real-time identification directly from video streams. The orthogonal fusion mechanism enhances the robustness of feature representations by simultaneously capturing fine-grained local details and comprehensive global context, thereby improving the model’s ability to handle occlusion and variations in appearance. To support rigorous evaluation, we construct a new dataset, Fishdata, containing 3000 annotated images collected under varied conditions. Experimental results demonstrate that our method consistently achieves over 97.5% recognition accuracy and maintains reliable performance even when fish targets are partially occluded or subject to appearance distortions. Overall, this work offers an effective, scalable, and practical solution for high-precision fish monitoring, contributing to the advancement of intelligent aquaculture systems and sustainable fisheries management.