MPR-net: Medicinal plant recognition network with dual-branch attention fusion

The recognition of medicinal plant images possesses substantial practical significance. However, there are very few methods for recognizing medicinal plants worldwide. Additionally, existing public datasets for medicinal plants are limited in both size and diversity, encompassing only a small number of species, which constrains the development of high-precision recognition models. To address these challenges, we have constructed a comprehensive medicinal plant image benchmark, named the Karst Landform Herbs Dataset, for medicinal plant recognition research and also proposed a novel Dual-Branch Attention Fusion-based Medicinal Plant Recognition Network, called MPR-net, for Medicinal Plant image recognition. Our constructed new Karst Landform Herbs Dataset contains 56,650 images representing 120 species of medicinal plants. To guarantee the effectiveness on natural images with complex background, the proposed MPR-net designs a dual-branch fused attention module for discriminative feature extraction, thereby enhancing its suitability for medicinal plant recognition tasks. Through a series of tests and evaluations conducted on the Karst Landform Herbs Dataset, as well as other publicly available medicinal plant datasets, experimental results demonstrate that the proposed method achieves outstanding performance in medicinal plant recognition, significantly enhancing recognition accuracy and model robustness, and maintains lightweight. Our code is available at https://github.com/lingf5877/MPR-Net, dataset is available at https://data.mendeley.com/drafts/pskvtbsmzw.