Multi-Scale Venation Pattern Analysis for Medicinal Plant Species Recognition

This research addresses the challenge of medicinal plant species recognition based on leaf images by focusing on venation patterns as discriminative features. Venation patterns - defined by the hierarchical arrangement of veins within a leaf - carry significant taxonomic information that is often overlooked by conventional plant classification approaches. We propose a novel, venation-aware methodology that combines specialized image preprocessing techniques with both transfer learning and custom-designed deep learning architectures. Our method extracts and analyzes venation patterns at multiple spatial scales, capturing both global and fine-grained structural details to improve classification performance. To validate the effectiveness of our approach, we developed and evaluated three distinct model architectures: 1) a modified ResNet-50 model utilizing transfer learning with an adapted input pipeline for venation-aware channels; 2) a custom-built convolutional neural network, VenationNet, explicitly designed for multi-scale venation analysis; and 3) a Dual-Stream CNN architecture that processes leaf texture and venation maps independently before merging via attention-based fusion. Preprocessing involves contrast enhancement, Frangi filtering for venation extraction, and edge detection to create a three-channel input comprising RGB, venation, and edge maps. Experimental evaluation using the Indian Medicinal Plants Dataset demonstrates that our venation-centric strategy significantly outperforms traditional CNN-based approaches, achieving higher accuracy, precision, recall, and F1-scores across diverse plant categories. This research contributes a practical and scalable solution for reliable medicinal plant identification, which is crucial for pharmacological research, biodiversity monitoring, and traditional medicine practices. Moreover, our approach is well-suited for deployment in real-time mobile and edge computing environments.