GAN Model Based Grape Leaf Disease Detection Using Deep Learning Algorithm with Avoid Overfitting Mitigation

Abstract

Timely detection of grape leaf diseases is essential for safeguarding vineyard productivity and preventing yield losses. This research presents a hybrid deep learning framework that combines generative data augmentation, transformer-based feature extraction, and capsule network classification to enhance multi-class disease recognition. The experiments utilized the Niphad Grape Leaf Disease Dataset (NGLD), containing 2,726 images categorized into Bacterial Rot, Downy Mildew, Healthy Leaves, and Powdery Mildew. To mitigate class imbalance, additional synthetic samples for the three disease classes were created using a Deep Convolutional Generative Adversarial Network (DCGAN), trained for 50 epochs per class. Real and generated images were integrated and processed through a pretrained Swin Transformer (Base, patch size 4, window size 7) to extract high-dimensional feature vectors, which were then reduced to 768 dimensions. These features were classified using an Attention-Guided Capsule Network (AGCapNet), enabling the model to focus on disease-specific patterns. The proposed method attained an overall accuracy of 96%, with a macro-averaged ROC-AUC of 0.97. A binary disease-versus-healthy analysis produced an AUC exceeding 0.95, highlighting the system’s suitability for early-stage disease identification. The results confirm that the integration of GAN-based augmentation with transformer and capsule architectures delivers a robust, scalable, and accurate approach to grape leaf disease detection.