Authors - Madhusmita Chakraborty, Vijay Kumar Nath, Deepika Hazarika Abstract - Due to morphological similarities between species, environmental variability, and the requirement for specialized knowledge, accurate identification of medicinal plants is still difficult, despite their critical role in primary healthcare systems around the world. A Hybrid RegNetX-Hierarchical Bidirectional Linear Cross Attention framework referred to as HR-HBCA framework for identifying medicinal plants from leaf photos is presented in this work. Multi-scale features are extracted using a RegNetX backbone, and computationally efficient linear crossattention is used in Hierarchical Bidirectional Linear Cross-Attentive Fusion (HBLCAF) blocks to integrate shallow spatial and deep semantic representations. Balanced contextual exchange across scales is achieved by bidirectional cross-attentive fusion. The HR-HBCA framework shows strong performance under notable intra-class variability, with accuracies ranging from 93.79% to 99.73% when tested on five diverse public datasets.
Open Zoom