A Novel Rar e Image Data Augmentation Method for Fish Diseases Classification

In the aquaculture business, timely monitoring and treatment of fish diseases can prevent substantial economic losses. Numerous advanced visual AI algorithms can recognize and classify fish diseases in time. However, collecting fish disease image data from natural environments remains costly, resulting in limited available data. This scarcity of fish disease images hinders the training of advanced visual AI models. In the paper, we propose a data augmentation method that uses image generation technology to address this challenge. Our method first collects healthy fish images and then employs object detection and image segmentation to identify fish body parts. We developed a random fish disease area mask generation algorithm that automatically identifies areas for disease feature generation. Finally, we use image generation models to create disease features in these masked areas, producing synthetic fish disease images. Our experimental results show that our proposed augmentation method improved the RestNet50 model's performance in fish disease image classification, improving precision by 6.0%, recall by 4.4%, F1-score by 3.8%, and accuracy by 4.4%. These improvements are superior to other common augmentation methods.