TSCG-TSM: a temporal-spatial-channel-global enhanced TSM model for fish feeding intensity assessment in recirculating aquaculture systems

Abstract

Accurate assessment of fish feeding intensity is essential for precision feeding in recirculating aquaculture systems (RAS). However, existing image-based methods cannot effectively capture temporal changes in feeding behavior, while conventional video-based models often introduce high computational complexity, limiting their practical deployment in real-time automatic feeding systems. Therefore, this study proposed a lightweight video classification model, temporal-spatial-channel-global enhanced TSM (TSCG-TSM), for fish feeding intensity assessment in RAS. The model adopted the Temporal Shift Module to enable 2D CNNs to extract spatiotemporal features through parameter-free temporal information exchange. On this basis, TSCG-TSM implemented three task-specific structural optimizations tailored to the fish feeding intensity assessment. First, the Temporal Attention Module was designed to replace the average consensus, allowing the model to adaptively assign weights to different temporal segments to enhance key-frame modeling. Then, the spatial-channel gated module enhanced feeding-related regions and semantic channels by recalibrating convolutional features across the spatial and channel dimensions. Finally, global response normalization can regulate global channel responses, enhance inter-channel competition, and improve feature diversity. Experimental results demonstrated that TSCG-TSM achieved an accuracy of 86.30% ± 1.94% and F1-score of 86.56% ± 1.98%. These are 4.11% and 4.35% higher than baseline, respectively, while increasing Params and FLOPs by only 0.87 M and 0.85 G. Moreover, TSCG-TSM achieved the best performance compared with 12 representative video-based models (best accuracy = 81.85% ± 2.23%, best F1-score = 82.15% ± 2.23%) These results demonstrate that TSCG-TSM achieves a balance between performance and computational complexity, showing potential for practical automatic feeding applications in RAS.