Efficient Neural Network Model Training for Fish Species Classification Using Attention-Enhanced MobileNetV3 in Aquaculture Applications

Abstract

Accurate fish species classification is a critical requirement in intelligent aquaculture management, yet existing deep learning models struggle to simultaneously achieve high accuracy and computational efficiency under resource-constrained conditions. This paper proposes a lightweight attention-integrated framework combining MobileNetV3 with the Convolutional Block Attention Module (CBAM). Two variants are developed: MobileNetV3-Large+CBAM achieves 98.98% accuracy, 98.90% F1-score, and 99.66% Top-3 Accuracy at 17.64 MB, while MobileNetV3-Small+CBAM offers an ultra-lightweight alternative with 98.01% accuracy at 6.50 MB and 1.59 GFLOPs. Both CBAM-augmented variants consistently outperform their baselines while introducing less than 3% additional FLOPs and model size. Experiments on a 31-species dataset of 13,304 images confirm the effectiveness and practical viability of the proposed framework for intelligent aquaculture monitoring, with broad applicability across automated surveillance, quality control, biological research, and education.