Fresh Fish Disease Detection in Aquaculture: Leveraging CNN Models for Accurate Diagnosis

Abstract

Aquatic products such as cultivating freshwater fish are indispensable for meeting the world’s protein demands. It becomes imperative to raise the fish production. But the industry faces a significant obstacle in the form of disease outbreaks among aquatic populations. These outbreaks not only lead to significant financial losses but also give rise to severe ecological concerns. The identification of diseased fish in aquaculture is still difficult due to lack of necessary infrastructure. It is crucial to identify contaminated fish at an early stage to stop the disease from spreading. Thus, a study on fish disease detection using CNN is proposed. A sample of about 3,000 fish images was gathered from open-source repositories with seven classes, including six typical diseases of freshwater fish and healthy fish. Resizing, noise reduction, normalization, and data augmentation of the images were performed beforehand to improve the generalization of the model. Multi-class disease classification VGG16, VGG19, InceptionV3, DenseNet, MobileNet and ViT, are six pretrained CNN architectures that were implemented and tested. Experimental findings show that DenseNet and MobileNet were the most precise, with the greatest classification accuracy of 90% and the highest recall as well as F1-score. This suggests that the CNN performs a reliable job of identifying and categorizing the type of diseases among infected fishes and provide a dependable and scalable solution to automated disease surveillance in aquaculture Furthermore, its focus on real-world image conditions, including noise and quality variations in photographs captured directly from aquaculture environments, makes it more suitable for practical deployment and automated field-level disease monitoring.