Enhanced U-Net and PSO-Optimized ANFIS for Classifying Fish Diseases in Underwater Images

Abstract

Fish diseases are among the major limiting factors to increase global aquaculture production. They lead to increased fish mortality, low breeding and growth rates, and low meat quality. The success of aquaculture is heavily dependent on the timely identification of disease. Therefore, we propose a fuzzy U-Net model to automatically identify fish disease from underwater images. U-Net is enhanced with multi-head channel and spatial attention and used to segment infected fish regions from fish disease images. Color pixel intensity features are then extracted from the localized regions, a form of guided feature extraction. Fuzzy C-means clustering is then used to find the cluster centroids and data distribution within the clusters, for the design of fuzzy membership functions. Moreover, the number of clusters are determined by silhouette score. Adaptive neuro-fuzzy inference system (ANFIS) is then trained, tested, and cross-validated for fish disease identification. The model parameters are optimized using particle swarm optimization (PSO) algorithm and compared with gradient-based methods. For image segmentation, the enhanced U-Net achieved a mean intersection over union (IoU) of 86.29%, mean pixel accuracy of 90.94%, mean precision of 93.58%, and mean recall value of 89.94% on 42 test images. Subsequently, ANFIS with PSO achieved overall superior performance on fish disease identification over gradient-based methods, with accuracy of 99.31%, precision of 99.00%, recall of 99.00%, and F1-score of 99.00%. The high-performance results of the optimized ANFIS confirm the robustness and efficacy of the proposed method to automatically identify fish diseases in aquaculture.