Synthetic Image Generation for Crop Disease Classification Using Generative Adversarial Networks

Due to biological diversity and unstructured surroundings, agricultural image analysis strives for optimal model performance to better accomplish visual identification objectives. Large-scale, balanced, and ground-truthed image datasets are very helpful, but they are frequently hard to come by, which restricts the creation of very effective models. The identification of plant diseases has benefited enormously from the continuous advancement of deep learning (DL) techniques, which provide a robust tool with incredibly accurate results. However, the efficiency of deep learning models is dependent on the quantity and caliber of labeled data used for training. Precise classification of crop diseases is important for precision agriculture. These models suffer from limited and imbalance datasets especially for rare diseases. The study suggests a framework using Generative Adversarial Network (GAN) for image generation to enhance the classification of diseases. The study employs conditional GAN trained on a Plant Village and New plant diseases datasets to generate synthetic images of diseased leaves. The images are evaluated using Structural similarity index (SSIM). Then the augmented images are integrated with the CNN classifier to measure the accuracy of disease prediction using synthetic dataset to validate the efficiency of image generation.