An investigative study on the usage of various GAN models for Plant Leaf Disease Detection

Plant leaf diseases may have a significant concern on the agricultural industry, and leads to substantial production loss. Early identification of plant leaf diseases in low-resolution images is a challenging task for improving crop yield, quality, and quantity of production. To address this challenge, a comprehensive review carried on various Generative Adversarial Network (GAN) models for the reconstruction of synthetic images. Employed a range of metrics, including a disclosure risk metric, to compare data generated by different GAN models. The same datasets are used to train the models, and the results were validated through generator loss, discriminator loss, and Universal Quality Index (UQI) scores. Our observations and discussions clearly state that the A-ESRGAN model with an attention U-Net structure as the discriminator generated more perceptually realistic high-resolution images than other GAN models. Subsequently, an Efficient-Net model trained on synthesized images to carryout robust understanding of disease patterns and characteristics for classification of plant leaf diseases.