Grape Leaf Disease Detection through Hierarchical Features, Mask R-Convolutional Neural Network Segmentation, and Self-Attention Networks
Automated grape leaf disease detection is essential for precision agriculture, enabling early diagnosis and effective disease management. This paper presents a novel deep learning-based framework to enhance the accuracy and robustness of grape disease detection by integrating advanced image pre-processing, segmentation, and feature extraction techniques. The proposed model addresses key challenges in traditional approaches by incorporating deep learning-based noise filtering, Mask R-CNN for instance segmentation, and a hybrid feature extraction strategy combining Residual Network (ResNet) and handcrafted methods such as Gray-Level Co-occurrence matrix (GLCM), shape and texture features. Additionally, an improved Mask R-CNN architecture is introduced by replacing the ResNet backbone with a Swin Transformer, leveraging hierarchical feature extraction and self-attention mechanisms for enhanced contextual understanding. Experimental results exhibit the efficacy of the proposed approach in accurately segmenting and classifying grape leaf diseases by reporting the classification accuracy as 92.75 and segmentation accuracy as 98.10%, respectively