CycleGAN-Based Data Augmentation with CNN and Vision Transformers (ViT) Models for Improved Maize Leaf Disease Classification
Syed Taha Yeasin Ramadan, Tanjim Sakib, Md. Ahsan Rahat, Shakil Mosharrof
Abstract
Crop losses pose a serious danger to global food security, and this problem also affects maize crops. To successfully address this issue, precise disease detection techniques are required. However, a major hurdle to developing reliable models to address this issue is the dearth of datasets. In re-sponse, we present a novel approach that uses synthetic images created by CycleGAN to supplement constrained datasets. We thoroughly assessed deep learning models, such as ResNet50V2, DenseNet169, VGG16, VGG19, Xception, MobileNetV2, and emerging vision transformer models, such as ViT-B/16 and ViT-B/32, with a focus on the two critical classes of maize leaf diseases, blight and common rust. Notably, DenseNet169 performed better than other models with an accuracy of 98.48%, especially when trained on the CycleGAN -enhanced dataset. CycleGAN-augmented data outperformed the performance of the models trained solely on the original dataset, demonstrating the effectiveness of the augmentation approach in performance enhancement. By utilizing CycleGAN's synthetic images, this study expands the field of maize leaf disease diagnosis and establishes DenseNet169 as a viable model for precise disease identification. The findings of the study have the potential to significantly revolutionize agricultural operations using advanced maize disease detection techniques.