Enhancing Potato Leaf Disease Detection: Implementation of Convolutional Vision Transformers with Synthetic Datasets from Stable Diffusion

Numerous studies have addressed the classification of potato plants. However, the available datasets often lack the necessary diversity to improve the accuracy of predictive classification models effectively. Our research capitalizes on synthetic datasets generated through the Stable Diffusion 1.5 image generation method to address this challenge. This study suggests a new way to solve the problem by using artificial datasets created with the Stable Diffusion 1.5 method to teach a Convolutional Vision Transformer (CvT) model how to identify diseases on potato leaves accurately. Our objective is to train the CvT model employing synthetic datasets to excel in detecting potato leaf diseases. Our methodology encompasses the model's training using synthetic datasets from Stable Diffusion 1.5. We employ a comprehensive dataset of 11,121 synthetic images to train the Convolutional Vision Transformer (CvT) model, which enables it to accurately identify various potato leaf diseases such as black leg/soft rot, mosaic, leaf roll, early blight, and late blight. We conduct evaluations at multiple training stages to gauge the model's performance and accuracy. The outcomes of our research underscore the effectiveness of employing synthetic datasets from Stable Diffusion 1.5, which significantly augments the available image data while preserving a high level of accuracy. The CvT model proficiently identifies potato leaf diseases with an evaluation accuracy of 84%. Additional testing reveals that by the fifth epoch, the CvT model attains an accuracy of 81% when assessed using 82 randomly selected images of diseased plants from Google. The implications of this research are far-reaching, particularly within the domains of image processing and agriculture. The strategy of utilizing synthetic datasets to train the CvT model presents an efficient remedy to address the limitations of original image datasets. The adept disease detection capability of the CvT model holds the potential to expedite plant condition identification, mitigate crop loss, and ultimately amplify agricultural productivity. This study effectively demonstrates that the Convolutional Vision Transformer (CvT), when leveraged with synthetic datasets from Stable Diffusion 1.5, produces a model capable of accurately identifying potato leaf diseases. These findings bear positive implications for both the agricultural and image-processing sectors.