A segmentation-combination data augmentation strategy and dual attention mechanism for accurate Chinese herbal medicine microscopic identification.

Abstract

Introduction: Chinese Herbal Medicine (CHM), with its deep-rooted history and increasing global recognition, encounters significant challenges in automation for microscopic identification. These challenges stem from limitations in traditional microscopic methods, scarcity of publicly accessible datasets, imbalanced class distributions, and issues with small, unevenly distributed, incomplete, or blurred features in microscopic images.

Methods: To address these challenges, this study proposes a novel deep learning-based approach for Chinese Herbal Medicine Microscopic Identification (CHMMI). A segmentation-combination data augmentation strategy is employed to expand and balance datasets, capturing comprehensive feature sets. Additionally, a shallow-deep dual attention module enhances the model's ability to focus on relevant features across different layers. Multi-scale inference is integrated to process features at various scales effectively, improving the accuracy of object detection and identification.

Results: The CHMMI approach achieved an Average Precision (AP) of 0.841, a mean Average Precision at IoU=.50 (mAP@.5) of 0.887, a mean Average Precision at IoU from .50 to .95 (mAP@.5:.95) of 0.551, and a Matthews Correlation Coefficient of 0.898. These results demonstrate superior performance compared to state-of-the-art methods, including YOLOv5, SSD, Faster R-CNN, and ResNet.

Discussion: The proposed CHMMI approach addresses key limitations of traditional methods, offering a robust solution for automating CHM microscopic identification. Its high accuracy and effective feature processing capabilities underscore its potential to modernize and support the growth of the CHM industry.