FewMedical-XJAU: A Challenging Benchmark for Fine-Grained Medicinal Plant Classification

Area of Science:

• Computer Science
• Botany
• Machine Learning

Background:

• Existing fine-grained plant image classification (FPIC) datasets lack diversity and real-world complexity.
• Challenges include limited categories, uniform backgrounds, and insufficient environmental variations.
• These limitations hinder the effectiveness of FPIC in practical applications.

Purpose of the Study:

• To introduce the FewMedical-XJAU dataset for rare medicinal plants in Xinjiang, China.
• To develop an improved FPIC method (BDCC) leveraging textual information for enhanced discrimination.
• To address challenges of high intra-class variability and inter-class similarity in plant classification.

Main Methods:

• Developed the FewMedical-XJAU dataset with diverse backgrounds, lighting, and expert annotations.
• Proposed Bilinear Deep Cross-modal Composition (BDCC) integrating textual priors into deep metric learning.
• Implemented Class-Aware Structured Text Prompt Construction and a dynamic fusion mechanism.

Main Results:

• The FewMedical-XJAU dataset provides a realistic testbed for FPIC.
• BDCC demonstrated superior accuracy and robustness in few-shot classification experiments.
• The proposed method effectively handles complex environmental conditions and plant variations.

Conclusions:

• The FewMedical-XJAU dataset and BDCC method significantly advance fine-grained plant image classification.
• This work offers robust solutions for identifying rare medicinal plants in challenging natural settings.
• The findings support the practical application of advanced image classification techniques in botany and conservation.