基于合成图像数据集和深度学习的中药材识别方法

Chinese Herbal Medicine Recognition Method Based on Composite Image Dataset and Deep LearningHUANG Jing (Xiamen Institute for Food and Drug Quality Control, Xiamen 361012, China)

关键词:  目标检测  中药材识别  图像合成  YOLOv5  对比学习

Key words:  target detection  Chinese herbal medicine recognition  image composition  YOLOv5  contrastive learning

摘要:

目的 设计中药饮片图像的自动合成方法和用于中药饮片识别的深度学习模型，实现对中药饮片的自动识别。方法 利用图像处理技术分割单个中药饮片样本的图像，而后随机抽取多个品种的中药饮片样本图像进行合成，生成训练数据集；并将对比学习技术与YOLOv5模型相结合，促使骨干网络提取更有效的特征，用于中药饮片的品种识别。结果 在合成和真实图像上的测试结果表明，对30个品种中药饮片的平均识别率超过了95%，平均每张图像的处理时间约为28 ms。边界框的精度和相似品种的识别率都得到了有效提升。结论 利用合成图像数据集训练的深度学习模型可实现对多个品种中药饮片的有效识别，并可推广到更广泛的中药饮片识别应用中。

Abstract:

OBJECTIVE To design an automatic composition method for Chinese herbal medicine images and a deep learning model to achieve automatic recognition of Chinese herbal medicine. METHODS Single sample image was segmented first and then randomly selected sample images of multiple types were synthesized to generate a training dataset. Contrastive learning was combined with YOLOv5 to enable the backbone to extract more effective features. RESULTS Experiment on composite and real images showed that the average recognition rate exceeded 95% and the average processing time per image was about 28 ms. The accuracy of bounding boxes and the recognition rate of similar types have been effectively improved. CONCLUSION The proposed method can effectively identify a variety of Chinese herbal medicine and can be extended to a wider range of Chinese herbal medicine recognition.