| Traditional Chinese medicine decoction pieces are widely used in the prevention,treatment and rehabilitation of diseases because of their advantages of regulating the whole body and having less side effects.However,there are so many types of traditional Chinese medicine decoction pieces that it is impossible for non-professionals to identify them.Even relying on professional means will consume a lot of manpower,material and financial resources,and the identification results are easily affected by subjective.Based on the fact that deep learning technology has been widely applied to image recognition in many fields and has achieved excellent results,the Convolutional Neural Network is expected to exert its superior performance in the recognition of traditional Chinese medicine decoction pieces.The research object of this paper is the image of decoction pieces of traditional Chinese medicine.In order to improve the recognition accuracy of the network model on the decoction pieces,the residual network will be improved;In order to realize the successful deployment of the network model to mobile devices with limited resources,the lightweight network model will be improved.The specific content is as follows:(1)Since there is no public and authoritative data set in the research field of traditional Chinese medicine decoction pieces,a small data set of 17 common traditional Chinese medicine decoction pieces was constructed by shooting to provide data support for subsequent model research on image recognition.(2)The recognition performance of five classic residual network models in the field of traditional Chinese medicine decoction pieces was discussed.The research results show that the recognition accuracy of Res Net34 on the test set is higher than that of the other four residual network models.In order to improve the accuracy of the model for image recognition of traditional Chinese medicine decoction pieces,the first convolutional layer with a convolutional kernel of 7 in the Res Net34 network model was decomposed and replaced by three serially connected convolutional layers with a convolutional kernel of 3,and added SE module.Propose SE-Res Net36 network model.The image recognition accuracy of this model is high,which can meet the requirements of high image recognition ability of traditional Chinese medicine decoction pieces.(3)In order to solve the problem that the classic Convolutional Neural Network has a large parameter memory,which cannot meet the needs of mobile devices with limited resources for the identification of traditional Chinese medicine decoction pieces,and the recognition accuracy of traditional Chinese medicine decoction pieces by the lightweight network is low,this paper proposes a multi-scale lightweight network model LW-Mobile Net V2(Light Weight Mobile Net V2).Based on Mobile Net V2,group convolution is used to construct a multi-scale feature extraction layer,and the model is compressed,and the h-swish activation function is introduced to improve the overall recognition ability of the model.The experimental results show that LW-Mobile Net V2 is significantly better than other lightweight network models in terms of recognition accuracy,parameter memory,and running speed.The model has high recognition accuracy,low calculation cost,small storage overhead,and strong practicability,which can meet the needs of mobile devices with limited resources for the recognition of traditional Chinese medicine decoction pieces.The above research work has achieved good results.Although the parameter memory of the SE-Res Net36 network model is 81.48 MB,the accuracy rate of image recognition of traditional Chinese medicine decoction pieces can reach 95.77%,which has a high recognition ability.The LW-Mobile Net V2 model recognition accuracy rate is 93.14%,while the parameter memory is only 4.85 MB,which is 43.4% smaller than Mobile Net V2,and has a high cost performance in terms of performance and storage costs. |
PDF Full Text Request