Gamma-Retinex vision transformer framework for automated grape disease detection

Abstract

Grape leaf disease poses a significant threat to crop yield and quality, demanding accurate and early detection for effective management. However, existing detection techniques face challenges such as overlapping color intensities, inconsistent lighting, excessive noise from preprocessing, and weak lesion boundary localization, all of which hinder classification performance. To overcome these limitations, this research proposes a deep learning framework, Gamma-snake Transformer Network (GST-Net), tailored for robust grape leaf disease detection. Unlike conventional approaches, the proposed framework integrates illumination-aware preprocessing, hybrid segmentation, and transformer-based feature learning to enhance disease discriminability under challenging environmental conditions. The model introduces a multi-stage pipeline where Gamma-Retinex Guided Filtering enhances image quality by correcting illumination artifacts and preserving structural edges. A GrabCut-Snake segmentation mechanism accurately isolates lesion regions by combining coarse initialization and refined contour evolution. Disease features are extracted using Vision Transformers (ViT), leveraging self-attention to capture subtle texture and color variations effectively. For classification, the system employs a hybrid approach: a fully connected Softmax classifier for known diseases, a Siamese Network for few-shot detection of novel classes, and a One-Class SVM to further validate unfamiliar patterns. Evaluated on a 4-class grape leaf dataset comprising 4062 images using an 80:20 train-test split and 5-fold cross-validation, the proposed GST-Net achieved an accuracy of 99.88%, precision of 99.89%, recall of 99.89%, F1-score of 99.89%, and specificity of 99.96%. Additionally, the model achieved an accuracy of 96.85%, precision of 95.90%, recall of 96.40%, and F1-score of 96.14% in detecting unseen disease classes, demonstrating its capability to identify unfamiliar disease patterns.