Development and Research of Real-Time Fish Passage Monitoring System for Fishway Based on YOLO-Fish Model

To address issues in real-time fish monitoring, this study proposes a fish detection model based on YOLOv11, termed YOLO-Fish. The model focuses on optimizing the core feature extraction module C3k2 and reconstructing it with lightweight backbone networks. Through a comparative analysis of MobileNet, GhostNet, and FasterNet, FasterNet is selected as the backbone network due to its ability to significantly reduce model parameters and computational complexity (FLOPs) while effectively preserving critical feature representations. Moreover, it achieves superior inference speed, making it well-suited for real-time fish passage monitoring scenarios. To validate the effectiveness and practicality of the proposed method, a fish passage experiment closely resembling real-world conditions was designed. The experiment was conducted on an indoor U-shaped flume platform at Zhejiang University of Water Resources and Electric Power, simulating typical scenarios such as multi-fish movement and flow disturbances. Four representative fish species from the Yangtze River basin—crucian carp, yellow catfish, grass carp, and Chinese sucker—were selected as detection targets. Based on this setup, a fish dataset was constructed using images captured by a GoPro HERO9 camera. Data augmentation techniques, including motion blur, elastic transformation, and illumination adjustment, were applied to enhance sample diversity. A semi-supervised learning strategy was then employed to annotate 3000 images with bounding boxes and class labels, which were converted into the YOLO format for training and validation. Experimental results demonstrate that, compared with mainstream models such as YOLOv10, YOLOv11, and YOLOv12, the proposed YOLO-Fish model achieves superior performance. On the constructed dataset, it attains an precision of 94.7% with an inference speed of up to 115 FPS. The proposed method exhibits excellent accuracy and efficiency in real-time fish monitoring under complex aquatic environments, providing a reliable technical solution for fish detection in ecological hydraulic engineering.