A Framework for Augmentation of Image Data using Transfer Learning Methods

Abstract

Accurate detection of plant leaf diseases plays a critical role in plant health management. Traditionally, plant pathologists relied on manual visual inspections and their expertise in plant diseases for identification, but this method proved subjective and error-prone. However, advancements in Machine Learning (ML) have brought about a transformative shift in disease identification. Modern systems automatically analyse leaf images and categorize them into distinct disease classes. This thesis presents a cohesive exploration of advanced methodologies, ranging from general data augmentation and transfer learning to specialized uncertainty-based and optimization techniques. This thesis research begins by addressing the limitations of existing techniques and delves into the synergy between data augmentation and transfer learning. The approach uses a Support Vector Machine (SVM) in conjunction with Transfer Learning (TL) techniques to extract features from pretrained layers and on evaluation of the approach achieved higher accuracy compared to current methods, specifically Deep Neural Networks (DNNs) and standard convolutional networks. Our results show an approximate 4-8% increase in accuracy over the previous methods. newlineNarrowing the focus to plant leaf disease classification, the thesis introduces the Uncertainty-Based Progressive Conditional Generative Adversarial Network (UPC-GAN) to mitigate overfitting risks. This approach generates augmented synthetic images, enriching the dataset and enhancing disease classification accuracy. Notably, UPC-GAN effectively captured image uncertainty through pixel-wise residual distribution, employing the Generalized Gaussian Distribution (GGD). Consequently, image augmentation using UPC-GAN improved the classification performance in the categorization of diseased leaves. newlineFurther the research specialization occurs with the integration of Differential Evolution-based Exponential Decay Learning Optimal Rate (DE-EDLOR) optimization and Transfer Learning using GoogLeNet (TLGN). EDLOR