Rapid identification of medicinal plants via visual feature-based deep learning

Area of Science:

• Computational botany
• Image recognition technology
• Traditional Chinese Medicine

Background:

• Traditional Chinese Medicinal Plants (CMPs) are integral to healthcare and cultural heritage in China, with a history spanning millennia.
• Accurate identification of CMPs is vital to ensure clinical safety and treatment efficacy, preventing confusion due to processing and cultivation variations.

Purpose of the Study:

• To develop an advanced method for accurate visual identification of Traditional Chinese Medicinal Plants (CMPs).
• To address challenges in CMP identification posed by data imbalance and feature representation.

Main Methods:

• Utilized a self-developed device for high-resolution data acquisition and constructed a visual multi-varieties CMPs image dataset.
• Implemented a random local data enhancement preprocessing method (random cropping and shadowing) for imbalanced data.
• Developed a novel hybrid supervised pre-training network integrating global features with Masked Autoencoders (MAE) via a parallel classification branch.
• Introduced newly designed losses based on reconstruction and classification to enhance training efficiency and learning capacity.

Main Results:

• The proposed hybrid network effectively captures features by integrating global information and local details.
• The novel data enhancement and loss functions improved training efficiency and learning capacity.
• Experimental results demonstrated superior performance compared to state-of-the-art methods on both custom and public datasets.

Conclusions:

• The developed method achieves state-of-the-art performance in CMP identification.
• The approach shows significant advantages in efficient implementation of plant identification technology.
• The findings indicate strong potential for real-world applications in the field of Traditional Chinese Medicine.