The Implementation Of Light Convolutional Neural Network In Mobile Terminal And Its Application In Heart Sound Prediction

Convolutional neural network,as a deep learning method,replaces the traditional manual feature selection by machine self-learning.In the fields of speech recognition,image classification and target detection,convolutional neural network has achieved achievements similar to or even beyond human visual recognition,and is still in the process of continuous development.Compared with the way of deploying CNN model to the server,deploying CNN directly on the mobile terminal has the advantages of zero delay,network independence and user privacy protection.However,the traditional convolutional neural network model is difficult to deploy to the mobile terminal due to its large size and high computational complexity.In view of the problems existing in the process of transplanting CNN to mobile terminal,this paper discusses the design of lightweight network,further compression and how to preprocess the data in the mobile terminal.Different CNN models often have different input streams and different output types.Considering these problems,this paper designs a system for the mobile terminal to acquire,preprocess and process the input data,and implements the task of heart sound prediction.In the task of heart sound prediction,tiny-VGG network designed for heart sound classification is used in this paper.Due to the large size of the model,in order to make the model suitable for the mobile end,we use the idea of deep separable convolution to modify the model,so that the model can take into account both the size and performance of the model.We use the heart sound data set on physionet and Pascal heart sound data set,after sorting out,we transform the heart sound data into two-dimensional timefrequency map,and use the modified CNN model for classification.The specificity of the model is 81.50% and sensitivity is 87.48% respectively.Compared with other CNN models,the results show that the modified model has good performance in size,accuracy and speed.Finally,to solve the problem of mobile deployment,we implemented CNN migration on Android mobile based on TensorFlow Mobile open source deep learning framework.The Android Package(APK)size of the heart sound recognition is 15.4Mb,and the single prediction time on the OPPO Reno Ace mobile phone is 62 ms.The terminal accuracy of the system is 75% on 20 measured data of the mobile system,and the specificity is 70% and the sensitivity is 80%.