DF-MobileNetV4: A dual-scale adaptive attention and fused ConvBlock enhanced MobileNetV4 for underwater fish feeding behavior recognition

Accurate assessment of fish feeding behavior enables aquaculture practitioners to optimize feed allocation and enhance production efficiency. While existing studies predominantly rely on surface-view recognition, this perspective suffers from water surface reflections, insufficient fish exposure, and limited camera angles. Underwater recognition effectively mitigates these constraints but faces two primary challenges: complex spatial distribution patterns during feeding (aggregation, dispersal, occlusion) that single-scale feature extraction cannot comprehensively capture, and subtle visual disparities between feeding intensities in underwater environments causing boundary ambiguity and misclassification susceptibility. To address these limitations, this study proposes DF-MobileNetV4, an improved MobileNetV4 incorporating dual-scale adaptive attention (DSA) and Fused ConvBlock. DSA jointly channel and spatial attention across multiple scales, achieving a 2.2% accuracy improvement over the baseline and demonstrating clear superiority over single-scale attention mechanisms. Fused ConvBlock enhances feature representation through dynamic channel expansion, reducing strong-to-medium boundary misclassification by 60% and contributing an additional 2.3% accuracy gain, effectively alleviating underwater turbidity-induced boundary ambiguity. In addition, we proposed RASUFI, a dataset comprising 4,944 underwater feeding intensity images from recirculating aquaculture systems. Comparative experiments against five mainstream models (e.g., ResNet34-CBAM and MobileViT-SENet) demonstrate that DF-MobileNetV4 achieves 95.59% recognition accuracy on RASUFI, representing a 2.7% improvement over baseline MobileNetV4, with only 4.04 M parameters, 1G FLOPs, and 15.4 MB model size. Cross-domain validation further confirms strong generalization capability, achieving 99.09% and 93.75% accuracy on underwater laboratory and surface-view datasets respectively, outperforming all comparison models on both benchmarks. Additionally, structured pruning improves accuracy to 96.49% while reducing model size to 12.5 MB, providing an efficient solution for intelligent feeding optimization in aquaculture applications.