PENERAPAN CONVOLUTIONAL VISION TRANSFORMER DALAM IDENTIFIKASI PENYAKIT DAUN KENTANG MENGGUNAKAN DATASET SINTETIS STABLE DIFFUSION

Various studies have conducted research on potato plant classification; however, the datasets often lack sufficient variation to enhance the accuracy of classification prediction models. This issue prompted the research to leverage synthetic datasets generated through the Stable Diffusion 1.5 image generation method. The proposed solution in this study is to utilize synthetic datasets generated through the Stable Diffusion 1.5 image generation method to train a Convolutional Vision Transformer (CvT) model for accurately detecting potato leaf diseases. The objective of this research is to train the Convolutional Vision Transformer (CvT) model using the synthetic dataset for the task of identifying potato leaf diseases. The methodology employed in this study involves training the model using synthetic datasets from Stable Diffusion 1.5. A total of 11,121 synthetic datasets are used to train the Convolutional Vision Transformer (CvT) model to identify potato leaf diseases such as black leg/soft rot, mosaic, leaf roll, early blight, and late blight. Evaluation is performed at various training stages to measure the model's performance and accuracy. The research results indicate that the use of synthetic datasets from Stable Diffusion 1.5 effectively expands the available image data while maintaining a high level of accuracy. The CvT model successfully recognizes potato leaf diseases with an evaluation accuracy of 84%. Further testing shows that by epoch 5, the CvT model achieves an accuracy rate of 81% when tested using 82 randomly selected images of diseased plants from Google. The implications of this research are highly relevant in the fields of image processing and agriculture. The approach of using synthetic datasets to train the CvT model provides an efficient solution to overcome the limitations of original image datasets. The accurate disease detection capability of the CvT model has the potential to expedite the process of identifying plant conditions, reduce crop loss, and overall enhance agricultural productivity. This study successfully demonstrates that the application of the Convolutional Vision Transformer (CvT) with the use of synthetic datasets from Stable Diffusion 1.5 can yield a model capable of identifying potato leaf diseases with high accuracy. These findings can offer positive implications for the agricultural and image processing sectors. Keywords: Convolutional Vision Transformer (CvT), potato leaf disease identification, synthetic dataset, Stable Diffusion 1.5, agriculture, image processing