Leaf Based Tomato Plant Disease Detection Using Generated Images from WGP-ESR GAN

Plant diseases impact agricultural products in terms of quantity and quality, ultimately affecting the income of farmers. Early detection of plant diseases is crucial for improving agricultural yields. In this paper, we are using deep learning techniques for detecting the diseases of tomato plants based on leaf images. Image gathering of diseased plant leaves from various plants leads to unbalanced disease classes. Such unbalanced disease classes are not suitable for the effective training of disease classifiers. So we are proposing a deep-learning model, WGP-ESR GAN(Wasserstein Generative Adversarial Network with Gradient Penalty and Enhanced Super Resolution GAN) to create a well-balanced dataset. The model consists of two stages. In the first stage, WGAN-GP generates 3x64x64 sized tomato plant leaf images, which are then inputted to ESRGAN in the second stage to produce realistic 3x256x256 sized images. Initially, the model was trained with only healthy tomato leaf images to get a pre-trained model, and then we continued the training process on a pre-trained model with unhealthy leaves. This process generates a well-balanced dataset of each class size is 5357. Disease classifiers (VGG16, ResNet50, and DenseNet121) are employed, and the Plant Village dataset is used for training and performance testing. The proposed model's generated images exhibit higher realism than state-of-the-art models, and the average classification accuracy of the balanced dataset generated by WGP-ESR (99.69 percent) is significantly improved than state-of-the-art models.