A RepVGG-Style Regional Token Transformer for Grape Leaf Disease Classification

The major contributions of this work are as follows. First, we propose a novel hybrid grape leaf disease classification framework that integrates a RepVGG-style multi-branch convolutional backbone with a Transformer encoder, enabling effective learning of both local disease characteristics and global contextual relationships. Second, we introduce a regional token learning strategy that transforms the final 7×7 convolutional feature map into 49 high-level semantic regional tokens, rather than directly tokenizing raw image patches, thereby providing more informative representations for Transformer-based feature learning. Third, multi-head self-attention is employed to capture long-range dependencies among lesions, venation patterns, disease boundaries, and surrounding leaf tissues, improving discrimination between visually similar grape leaf diseases. Finally, extensive experimental evaluations demonstrate the effectiveness of the proposed framework, while Grad-CAM visualizations verify that the model focuses on biologically meaningful disease regions, enhancing both classification performance and model interpretability.