An Effective Data Augmentation Based on Uncertainty Based Progressive Conditional Generative Adversarial Network for improving Plant Leaf Disease Classification

Abstract

Early discovery and precise classification of plant-lead diseases are used to handle their spread and enhance the overall yield and product quality. The deep learning approach attains significant achievements in image classification and recognition. Since the classification using deep learning mainly depend on large-scale dataset for preventing the overfitting issue. Image augmentation is required to be developed to eliminate the risk of overfitting during the classification. In this research, the deep learning model based augmentation namely uncertainty based progressive conditional generative adversarial network (UPC-GAN) is developed for improving the plant leaf disease classification. The UPCGAN is used to map the images from one domain to another domain in a paired manner and estimate the uncertainty with the created images. Moreover, UPCGAN performs pixel wise residual distribution using the independent distributed zero mean generalized Gaussian distribution (GGD). The progressive learning of UPCGAN increases the differences in the augmented synthetic images for improving the classification using DenseNet121. The dataset used to evaluate the proposed UPCGAN-DenseNet121 method is PlantVillage dataset. The performance of UPCGAN-DenseNet121 is analysed using accuracy, precision, recall and F1-score. Existing research such as deep convolutional GAN (DCGAN)-GoogleNet, conditional GAN (CGAN)-DenseNet121 and Fast wide and deep feature extraction block (WDBlock) based GAN namely FWDGAN are used to evaluate the UPCGAN-DenseNet121. The accuracy of UPCGAN-DenseNet121 for 10 classes is 98.2%, which is high when compared to the DCGAN-GoogleNet, CGAN-DenseNet121 and FWDGAN.