A convenient method for the accurate identification of Citri Reticulatae Pericarpium using image and multi-stream

Area of Science:

• Agricultural Science
• Computer Vision
• Food Science

Background:

• Citri Reticulatae Pericarpium (CRP) quality is influenced by origin and aging, necessitating accurate authentication.
• Adulteration of lower-grade CRP to mimic premium products poses a significant challenge.
• Existing vintage classification methods are costly and complex, limiting practical application.

Purpose of the Study:

• To develop a convenient and accurate method for identifying the region and vintage of Citri Reticulatae Pericarpium.
• To overcome the limitations of traditional authentication methods for CRP.
• To enable quality assurance and fair market valuation of CRP.

Main Methods:

• Utilized an object detection network to localize exocarp and albedo regions in CRP images.
• Employed a three-stream feature extractor to process whole images and specific patches, capturing complementary visual details.
• Integrated a channel-level feature interaction module for enhanced robustness and a meta-learning module for device adaptation.

Main Results:

• Achieved 95.5% accuracy for CRP identification using images captured by consumer-grade devices (iPhone).
• Demonstrated a relative accuracy improvement of over 34% compared to direct transfer methods for cross-device image analysis.
• Highlighted the effectiveness of meta-learning in adapting to varying image capture conditions and devices.

Conclusions:

• The proposed image and multi-stream method offers a scalable and accurate solution for Citri Reticulatae Pericarpium authentication.
• Meta-learning significantly enhances the method's adaptability and performance across different imaging devices.
• This approach provides a foundation for reliable quality control and market valuation in the CRP industry.