Wheat Leaf Disease Synthetic Image Generation from Limited Dataset Using GAN

For deep learning models to be trained effectively, a sufficiently sizable and varied dataset must be available. Data augmentation, which produces identical images from a small number of original training samples, has proven to be an effective strategy for addressing the problem of deep convolutional neural networks (DCNNs) missing sufficient training data. However, getting a good dataset is frequently challenging, particularly in the context of identifying plant diseases. In this article, we offer a distinctive approach for generating images of wheat leaf disease from a smaller dataset using generative adversarial networks (GANs). We investigate two well-known GAN architectures, DCGAN and CycleGAN, for producing wheat leaf disease images from a constrained number of real-world images. Our findings demonstrate that in terms of image quality and resemblance to real-world images, the CycleGAN architecture outperforms the DCGAN architecture. In different applications, such as the identification of plant diseases, our study shows the possibility of employing GANs to produce realistic images from fewer datasets.