| Chinese medicinal materials are the gems of traditional Chinese medicine culture,as well as the indispensable material carrier of traditional Chinese medicine.After thousands of years of accumulation and development,they play a vital role it until now.China boasts rich,diversified Chinese medical materials.Only experts can identify their kinds.The traditional identification methods rely mainly on experts’ feeling and experience,during which there exists strong subjective factors,low efficiency and other problems.Besides,the identification method based on traditional machine learning is complex in operation and low in efficiency.This hinders the popularization and development of traditional Chinese medicine.Therefore,this paper introduces the method of deep learning to study the identification methods of Chinese medicinal materials,and builds an image recognition model of Chinese medicinal materials based on Faster R-CNN convolutional neural network.On the basis of this model which will be upgraded,Web end common medicinal materials identification platform was developed in a way that expands the application of this technology in Chinese medicinal materials.The main work is as follows:(1)Given that there are no standard data sets in Chinese medicinal materials image recognition,image data of Chinese medicinal materials were collected through network crawling and actual shooting to establish commonly used Chinese medicinal materials image data sets including amomum,beartiful sweetgum fruit,brucea,dwarf lilyturf tuber,forsythia,fritillaria cirrhosa,fructus tsaoko,myristica,raspberry and semen pruni.The image data were screened,expanded and unified in format,size and naming.Label IMG labeling software was used to manually label the categories and positions of Chinese medicinal materials in the images to establish VOC format data set containing 6089 Chinese medicinal materials images.This meets the basic needs of model practice.(2)By comparing the experimental results of two classical target detection algorithms on the number image data set of Chinese medicinal materials,the Faster R-CNN convolutional neural network was selected to build the image recognition model of common Chinese medicinal materials.As a result,the recognition and classification of Chinese medicinal materials were realized.(3)To further improve the detection performance,the established Chinese medicinal materials image recognition model has been optimized.First of all,Res NET-50 with stronger feature extraction ability and fewer parameters was chosen as the new backbone feature extraction network.In addition,the feature fusion mechanism was introduced and the generation strategy of anchor frame was adjusted to solve the large scale difference of Chinese medicinal materials in images.What’s more,Soft-NMS non-maximum suppression algorithm was employed as a new target localization method to alleviate the missing detection of medicinal materials under overlapping conditions.In the end,the improved model was compared with the previous one,and the identification ability of the model was tested in various environments.The results show that the average detection accuracy m AP of the improved model increased to 96.57%,an 2.7% increase than the previous model.The identification effect of the model is improved obviously.(4)Web-based identification platform for common Chinese medicinal materials based on the Flask framework was built on the basis of improving Chinese medicinal materials recognition modle.The demand analysis and function design have been conducted,and the recognition methods of Chinese medicinal materials have been enriched.The experimental test shows that platform can quickly and accurately realize the classification and identification of ten kinds of common Chinese medicinal materials.This meets the demand of real-time detection and brings certain practical value. |
