Dual-path feature extraction with transformer classifier for leaf disease classification

Leaf diseases pose a significant threat to agricultural productivity. The artificial intelligence-based deep learning models have been widely explored for plant disease classification. But, they suffer from limitations of inadequate multi-scale feature extraction and weak structural representation. In this work, novel dual-path feature extraction is proposed for automatic leaf disease classification. The first feature extractor uses a Contextual Feature Encoder Network (CFE-Net) module for feature extraction. It uses asymmetric convolutional blocks and dilated convolutions to extract multi-scale contextual patterns from leaf images. The second extractor, GSA-Net (Gradient-Selective Attention Network), extracts low-level structural features by highlighting edges and textures using Laplacian filtering and convolutional operations. The extracted features from both networks are fused using a Cross-Channel Adaptive Fusion (CCAF) block, which dynamically selects and weights relevant features. Finally, the Transformer-based classifier is used to map the fused representation to class probabilities. The proposed experimental evaluations on PlantVillage show that it achieves an average accuracy of 96.9% compared to previously proposed models.