A Multi-Phase Entropy-Fuzzy Framework for Precision Segmentation of Ayurvedic Leaf Diseases by HAFES-Net

Abstract

The accurate segmentation of disease-affected areas in medicinal plant leaves is a critical step for early diagnosis, yield preservation, and quality assurance in herbal medicine production. Traditional methods such as manual inspection and basic image thresholding are often ineffective due to variations in lighting, complex leaf textures, and subtle disease symptoms. This paper introduces HAFES-Net (Hybrid Attention-Fuzzy Entropic Segmentation Network), a novel multi-phase segmentation framework designed for Ayurvedic medicinal plants such as betel, tulsi, basil, and gotu kola. HAFES-Net integrates entropy-guided lesion localization, adaptive fuzzy thresholding, vegetation index suppression, morphological refinement, SLIC superpixel clustering, and marker-controlled watershed segmentation to achieve highly accurate disease isolation. The framework operates on denoised inputs produced by an AD³ entropy-weighted anisotropic diffusion filter, ensuring preservation of fine lesion boundaries while reducing noise. Experimental evaluation against benchmark methods—Otsu Thresholding and K-means Clustering—demonstrates superior performance, with average sensitivity, specificity, Jaccard Coefficient (JC), Dice Coefficient (DC), and accuracy values of 96.84%, 97.52%, 93.11%, 96.48%, and 97.26%, respectively. The proposed approach delivers robust and interpretable segmentation results across diverse disease patterns and illumination conditions, making it suitable for field-deployable plant health monitoring systems.