Multi-scale Attention Lightweight Network for Underwater Schooling Fish Detection and Behavioral Analysis

With the deepening of global underwater ecological research, there is an increasing demand for underwater fish behaviour analysis and detection, especially for real time, automatic and accurate detection of underwater fish, which is of great scientific value and practical significance. In this study, we propose an innovative Multi-scale Attention Lightweight Network (MAL-Net) that focuses on underwater fish detection and behaviour analysis. With the powerful adaptive feature extraction and deep learning and attention mechanism, the study overcomes the difficulties of complex underwater environment and rapid fish movement to achieve efficient fish identification and behavioural pattern analysis. Conventional methods face many challenges, including low visibility, light variations, and dynamic backgrounds, which are addressed in this study by advanced deep learning techniques that provide higher detection accuracy and speed critical. The network also includes a velocity analysis component that provides valuable insights into fish movement patterns and behaviour, contributing to scientific research and resource management efforts. Experimented on a constructed challenging dataset demonstrate that MAL-Net exhibits superior detection accuracy and processing speed. The article systematically describes the architectural design of the network model, and analyses in detail the role and advantages of the multi-scale attention mechanism in improving the detection effect. On this basis, the superior performance of MAL-Net in terms of accuracy, speed and robustness is verified in combination with multi-scenario comparative experiments and in-depth analyses, and an outlook is provided for its future potential in a wider range of water ecosystem applications.