Accurate and real-time fish feeding intensity scoring using channel convolution GLU and SMFA attention fusion in recirculating aquaculture system

In aquaculture, real-time quantification of fish feeding intensity is critical for developing scientific feeding strategies. Previous deep learning-based studies primarily focused on rough recognition of feeding and non-feeding patterns. However, with the growing demand for intelligent feeding systems, there is an urgent need for precise and quantitative assessment of fish feeding intensity. To address this, this study proposes CS-TransNeXt, a fish feeding intensity scoring model that integrates Channel Convolution GLU (CCGLU) and Self-Modulation Feature Aggregation (SMFA), which can precisely quantify feeding intensity into 10 scoring scales (1−10). Specifically, the CCGLU module is introduced into the TransNeXt, so as to enhance local feature modeling by fusing multi-scale Depthwise Convolutions with channel attention. Meanwhile, the SMFA replaces the original multi-head self-attention in TransNeXt, enabling adaptive weight adjustment through variance-based dynamic parameters of global-local features. Experimental results demonstrate that the proposed CS-TransNeXt achieves a Top-1 Accuracy of 95.25 %, an F1-Score of 95.30 %, outperforming the baseline TransNeXt-micro by 4.00 % in accuracy. Meanwhile, it is only 17.60 M, and provides a novel method for high-precision quantitative scoring of fish feeding intensity, offering significant practical value for the development of intelligent feeding systems.