A Deep Learning Framework for Fish Disease Classification Using EfficientNetB7 and Gaussian-Inhibited Spatial Attention

Abstract

Accurate classification of fish diseases is essential for maintaining the health and productivity of aquaculture systems. This study proposes a novel deep learning framework that combines the EfficientNetB7 architecture with a custom attention mechanism called Neural Spatial Attention with Gaussian Inhibition (NSAG). The NSAG module integrates both channel and spatial attention pathways, enhanced by a Gaussian inhibition mechanism that sharpens focus on the most relevant spatial regions. To validate the effectiveness of the proposed approach, we performed extensive experiments on a dataset of fish disease images. Comparative evaluations were performed across five configurations: (1) fine-tuned EfficientNetB7, (2) EfficientNetB7 combined with Squeeze-and-Excitation (SE) attention, (3) EfficientNetB7 with spatial attention, (4) EfficientNetB7 integrated with Convolutional Block Attention Module (CBAM), and (5) our proposed EfficientNetB7 + NSAG architecture. All models were trained under identical conditions for fair comparison. The experimental results provide evidence that the proposed model consistently outperforms the competing baselines, achieving an accuracy exceeding 97.94%, with an improvement of nearly 3\% over the CBAM-based model. These findings demonstrate the potential of NSAG in boosting attention learning and improving the classification of fish diseases.