An attention-guided multi-scale feature cascade network for underwater fish counting

Abstract

Visual counting is essential for advancing fisheries intelligence, but fish scale variation in open underwater environments has made underwater fish counting a constant challenge. Therefore, we propose an Attention-guided Multi-scale Feature Cascade Network, named AMFCNet, which resolves scale variation and improves the accuracy of fish counting in complex underwater environments. AMFCNet utilizes a multi-scale attention gate for multi-scale feature fusion, and integrates a multi-scale convolution module to capture complex spatial relationships. It also employs a multi-head supervision fusion strategy to mask irrelevant regions, ensuring targeted learning for each scale and generating high-quality multi-scale density maps. Experimental results demonstrate that the proposed method achieves state-of-the-art performance on the proposed dataset with the lowest computational cost, significantly outperforming 11 mainstream counting methods. It also achieves excellent results on other publicly available underwater datasets, with Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Normalized Absolute Error (NAE) values of 1.26, 1.71, and 0.08, respectively. This method shows significant potential for practical applications in aquaculture, such as in marine ranching and pond farming, to assess fish growth conditions and adjust feeding strategies accordingly.