Fusing appearance and vein morphology using dual-branch deep networks for accurate medicinal plant identification

Accurate identification of medicinal plants from leaf images is essential for pharmacognosy, biodiversity conservation, and agricultural decisions. But, accurate identification of medicinal leaves still poses a potential challenge in real-world conditions due to high similarity between species, variability within classes, uneven lighting, background clutter, partial views and occlusions. Existing RGB-based deep models often overfit to color-texture cues that vary with environmental conditions, whereas venation-based (skeleton) methods provide anatomically stable morphology but inherently suppress the critical appearance information needed to distinguish visually similar species. In this study, we introduced a novel dual-branch deep learning framework that explicitly separates and preserves appearance and venation learning using two independent pre-trained feature extractors, instead of relying on traditional fusion methods that combine the modalities at the input level or compress both cues into a single fused image stream. Specifically, MobileNetV2 is used to capture global appearance descriptors (texture, pigmentation, and shape), while DenseNet121 learns fine-grained vascular topology from skeletonized vein representations; the resulting embeddings are then combined via late feature-level fusion to form a unified discriminative representation that minimizes modality interference and maximizes complementarity. To further improve robustness and reduce bias introduced by dataset imbalance, we have integrated a class-frequency aware augmentation strategy that adaptively strengthens minority-class transformations while preserving majority-class fidelity, alongside transfer learning, class weighting, and regularization. The proposed approach is trained and evaluated on a curated dataset of 14,344 paired RGB-skeleton images spanning seven medicinal plant species. It is rigorously benchmarked against RGB-only, skeleton-only, and fused image baselines. Experimental results have shown that the proposed dual-branch model achieves 97 % overall accuracy with high precision, recall, and F1-score, showcasing that the structured dual-stream learning of appearance and vein morphology provides a solution for medicinal plant recognition with the potential for robust performance in changing and real-world settings.