Intelligent Screening of Pieces of Chinese Medicine Based on BMFnet-WGAN
10.13422/j.cnki.syfjx.20210819
- VernacularTitle:基于BMFnet-WGAN的中药饮片智能甄别
- Author:
Yan CHEN
1
;
Li-si ZOU
2
Author Information
1. Medical College，Yangzhou Institute of Technology，Yangzhou 225000，China
2. College of Pharmacy，Nanjing University of Chinese Medicine，Nanjing 210000，China
- Publication Type:Research Article
- Keywords:
traditional Chinese medicine （TCM）;
pieces;
feature learning;
semantic description;
Wasserstein
- From:
Chinese Journal of Experimental Traditional Medical Formulae
2021;27(15):107-114
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the feasibility of machine vision and deep learning methods in intelligent screening of pieces of Chinese medicine， so as to meet the needs of modern screening of pieces of Chinese medicine and overcome the problems of strong subjectivity and low efficiency in traditional screening based on manual experience. Method:An image set containing 11 125 images for 60 kinds of pieces of Chinese medicine was constructed， and the network architectures for high- and low-frequency feature learning were designed. Specifically， the parallel convolutional network was employed to obtain the low frequency feature and the deep multi-scale convolutional neural network to uncover the high-frequency feature. The semantic network was used to realize the feature learning mode with generalization ability. In this study， Wasserstein distance was introduced into the generative adversarial networks （GANs） to complete the screening of pieces of Chinese medicine， and the conditional parameters were added to the generation and discrimination networks to make the network training more reliable and improve the accuracy of identification. Result:The experiment results showed that when the ratio of training samples to test samples was greater than 6∶4， the identification accuracy of pieces of Chinese medicine was relatively stable. The identification accuracy of images captured in different states and environments by bi-view multi-feature network Wasserstein generative adversarial network （BMFnet-WGAN） reached up to 85.9% on average and the stability was high， demonstrating that BMFnet-WGAN was superior to VGG-Net and AlexNet. Conclusion:The BMFnet-WGAN method enables the revealing of rich and typical characteristics of decoction pieces and the introduced WGAN model and Wasserstein distance make the network training more reliable. The resulting accuracy， robustness， and batch effects in the intelligent screening of pieces of Chinese medicine were good， which has provided the technical support for the sorting and quantitative quality screening of pieces of Chinese medicine.
