Pearl Millet Disease Detection with Transfer Learning and cGAN-Driven Data Augmentation

In this study, an advanced method to detect disease in Pearl Millet is proposed wherein Transfer Learning along with Conditional Generative Adversarial Networks (cGANs) are applied for data augmentation. For addressing dataset quality and imbalance problem, the work uses cGANs to generate realistic synthetic images of diseased millet leaves with latent dimensions of 50, 100, and 500. This augmented dataset of plant roots helps improve the classification performance as it provides more quantity and more diversity of training data. For the classification task, six fine-tuned models are used; these are: ResNet-50, ResNet-101, ResNet-152, InceptionResNet, InceptionV3, and VGG-19; optimizer is Stochastic Gradient Descent (SGD). The performance of these models was trained by 100 epochs and evaluated by accuracy, precision, recall, F1 score and ROC curve. However, ResNet-50 and ResNet-152 are among those and their remarkable accuracy of 99% is better than the previous work’s 98% accuracy, proving this integrated approach to disease detection can be effective. The results show that transferring knowledge from a prior set of training data, by applying transfer learning, improves disease classification accuracy, along with applying data augmentation, which allows the creation of additional data, and robust model evaluation, to ensure accurate evaluation.