Fish Abnormal Behavior Detection and Classification via Multi-feature Fusion and Multi-tiered Hypergraph Network

Abstract

Real-time monitoring of abnormal fish behaviors plays a crucial role in enhancing management efficiency, reducing disease risk, and optimizing feed strategies in modern intelligent aquaculture. To address challenges such as image blurring, complex backgrounds, behavioral diversity, and the trade-off between detection accuracy and computational efficiency, a robust detection model, YOLO 11-AB, was presented based on an improved YOLO 11 architecture. The model incorporated a multi-scale convolution module (C3k2_PKI Module) in the backbone network to enhance perception of behavioral features at various scales. A lightweight mixed local channel attention (MLCA) mechanism was integrated into the feature extraction stage to effectively fuse channel and spatial information, thereby improving feature representation. Additionally, the neck of the model adopted a hypergraph-based cross-level representation network (HyperC2Net) and a mixed aggregation network (MANet), further strengthening the detection and discrimination of abnormal behaviors in complex underwater environments. Experimental results demonstrated that the proposed model achieved significant improvements in detection accuracy, computational efficiency, and classification performance, with a 3.3 percentage points increase in precision, a 5.9 percentage points increase in recall, and a 4.4 percentage points increase in mean average precision compared with traditional methods. This approach provided technical support for early warning and management of fish diseases in high-density factory aquaculture, offering practical value for enhancing efficiency and promoting the sustainable development of the aquaculture industry.

Keywords: smart aquaculture, abnormal behavior detection, multi-scale feature fusion, attention mechanism, YOLO 11-AB model