AI-powered recognition of Chinese medicinal herbs with semantic structure modeling and gradient-guided enhancement

Jiaxing Zou ¹, Ruiwei Liao ¹

⁰Guangzhou University of Chinese Medicine, The Research Center of Basic Integrative Medicine, Guangzhou, China.

Frontiers in Plant Science +

|

November 24, 2025

View abstract on PubMed

Summary

This study introduces a novel AI framework for accurate medicinal plant recognition using image analysis. The structure-aware approach enhances identification accuracy in visual sensor applications.

Area Of Science

Computer Vision
Artificial Intelligence
Biotechnology

Background

Accurate identification of medicinal plants is crucial for drug discovery and traditional medicine.
Existing methods for herbal recognition face challenges with fine-grained details and complex visual data.

Purpose Of The Study

To develop a structure-aware artificial intelligence (AI) framework for fine-grained recognition of medicinal plant images.
To improve upon current herbal recognition techniques by integrating advanced AI modules.

Main Methods

The proposed framework utilizes a Swin-Transformer backbone, incorporating graph-based structural modeling, a Bidirectional Semantic Transformer, and a Gradient Optimization Module.
This approach captures spatial and channel-wise dependencies for enhanced feature discrimination.

Main Results

Achieved 90.32% accuracy on the TCMP-300 dataset (52,089 images, 300 categories), surpassing the Swin-Base baseline by 1.11%.
Obtained 92.75% accuracy on a self-constructed herbal dataset, outperforming baselines by 1.18% despite challenging conditions.

Conclusions

The developed framework offers robust and sensor-adaptable solutions for practical plant-based applications.
This AI-driven approach significantly advances the field of herbal recognition beyond existing models.

Keywords:

bidirectional transformer deep learning fine-grained image classification medicinal plant recognition semantic attention