Application of EfficientNet Deep Learning with Wiener Filter for Freshwater Fish Disease Image Classification

Abstract

Challenges pertaining to the timely and accurate diagnosis of diseases in freshwater fish have adversely impacted the productivity of the aquaculture industry. Image classification using deep learning techniques has the potential to overcome such challenges. However, this potential has not been realized due to such problems as image noise, motion blur, and small dataset sizes. Most prior studies in this area employ the same Convolutional Neural Network (CNN) architectures and, while using the same or similar techniques, are generic to the studies to preprocess the images. The focus of this study is to compare and benchmark the image classification performance of the EfficientNet architectures (B0 to B7) using the Wiener Filter as a preprocessing technique for the classification of diseases in freshwater fish. The experiments used a publicly available dataset of 1,750 images of seven diseases in fish while maintaining identical training parameters to yield sixteen different experimental configurations. Metrics such as accuracy, precision, recall, and F1-score were exercised while evaluating model performance. The data show that medium-scale architectures surpass both smaller and larger size variants. The optimal performance was achieved by EfficientNet-B4 and the Wiener Filter with an accuracy of 94.89%, a precision of 95.15%, a recall of 94.92%, and an F1-score of 94.89%. The results confirm that preprocessing with the Wiener filter improves performance on classification tasks using medium-sized models and further elucidate the applicable value of the model developed in this study in aquaculture and its related interventions.