A Plant Disease Identification and Localization Approach without Mask Labeling

In our research, we investigated an alternative approach for the identification and localization of plant diseases utilizing a system built on Generative Adversarial Networks (GANs) and autoencoders. The proposed system addresses the difficulties in identifying diseased leaves from limited training dataset by training solely on healthy plants. By coupling the GAN and encoder models the healthy version of input plant images can be generated. The localization of diseased areas is accomplished by comparing the actual image with their healthy counterparts, generated by the proposed system, using Mahalanobis distance, a metric that measures the extent of variation between the actual image and the generated image. Exceeding a certain threshold indicates leaf disease, pinpointing the exact area for disease identification. The proposed results, validated by comparing Intersection over Union (IoU) metrics with manual labeling, and getting the classification performance from the plant disease classification model, confirm that the system has a promising performance in accurately identifying and localizing leaf diseases.