Generation and Transformation Invariant Learning for Tomato Disease Classification

Abstract

Deep learning-based plant disease management became a cost-effective way to improved agro-productivity. Advanced train sample generation and augmentation methods enlarge train sample size and improve feature distribution but generation and augmentation introduced sample feature discrepancy due to the generation learning process and augmentation artificial bias. We proposed a generation and geometric transformation invariant feature learning method using Siamese networks with maximum mean discrepancy loss to minimize the feature distribution discrepancies coming from the generated and augmented samples. Through variational GAN and geometric transformation, we created four dataset settings to train the proposed approach. The abundant evaluation results on the PlantVillage tomato dataset demonstrated the effectiveness of the proposed approach for the ResNet50 Siamese networks in learning generation and transformation invariant features for plant disease classification.