YOLOv10-UDFishNet: detection of diseased Takifugu rubripes juveniles in turbid underwater environments

Abstract

Accurate detection of diseased Takifugu rubripes is essential for effective aquaculture disease management. However, turbid underwater conditions and the small scale of juvenile fish result in low precision. This paper proposes a diseased fish detection model called YOLOv10-UDFishNet, which comprehensively optimizes the backbone network, neck network, and detection head of YOLOv10. The model features an innovative backbone network with an auxiliary branch to capture multi-level image information, thereby enhancing feature extraction for diseased fish. Furthermore, space-to-depth convolution (SPD-Conv) replaces traditional strided convolution to preserve detail information, achieving accurate detection of diseased juveniles. In the neck network, we propose a feature fusion method based on high-level feature guidance (HFG), dynamically adjusting the concatenation of disparate feature maps to achieve higher recall. To sharpen the focus on diseased fish, the detection head utilizes the SlideLoss classification loss function, which improves detection performance. Ablation and comparison experiments are designed on a self-constructed dataset to verify the network’s superiority. The precision and recall of YOLOv10-UDFishNet were 94.0% and 90.1%, respectively, which are improved by 2.2% and 4.8%, respectively, over those of YOLOv10. Compared with current state-of-the-art underwater detection models such as DDEYOLOv9 and DFYOLO, our model achieves 2.5% and 3.1% relative improvements, respectively, in diseased fish recall. The results show that YOLOv10-UDFishNet is more adaptable to turbid underwater environments and exhibits enhanced effectiveness in diseased Takifugu rubripes detection.