A Data Augmentation Method Based on GAN for Plant Disease Recognition

Significant advancements have been achieved in plant disease recognition through deep learning techniques. However, these approaches often face overfitting due to the difficulty in acquiring large-scale, balanced, and authentic plant disease images. To address these challenges, this paper proposes a data augmentation method based on CycleGAN. It implements a leaf mask construction module by integrating MobileViT with Grad-CAM++, employ deep residual shrinkage network to suppress noise and utilize dual paths to fuse features and improve the CycleGAN generator. Finally, we integrate this module with improved CycleGAN to achieve the conversion of healthy leaf images to diseased leaf images. Experimental results show that using the images generated by this method to augment the plant disease data and train the recognition models, the recognition accuracy for seven soybean diseases reaches 92.33% on ResNet-50, surpassing the performance of alternative data augmentation techniques.