Research On Face Recognition Based On Sparse CNN On Microcontroller Chip

In recent years,with the improvement of the performance of face recognition algorithms,face recognition technology has been applied in many fields.At present,the face recognition technology on the market is mainly divided into two categories:verification technology(comparison between face and ID photo)and search recognition technology(face recognition based on database).This article mainly studies search recognition technology.Access control,attendance and other systems use various algorithms to directly perform face recognition on the special equipment based on the operating system.The cost of the device is high,the power consumption is large,and real-time and security cannot be guaranteed.In the fields of security and finance,a large number of images are collected and transmitted to the back-end server for data calculation and processing through front-end equipment,which not only increases the pressure on the server,but also consumes a lot of network bandwidth.Therefore,it is of great significance to use neural network algorithms to study face recognition on low-power,low-cost microcontrollers.A method for implementing face recognition based on a sparse neural network on a microcontroller chip is proposed aiming at the above problems.The face database is used to train the convolutional neural network on the server.Compress the model parameters using a sparse algorithm,fix the network model and export the parameters of each layer after the training is completed.CMSIS-NN library is used to build the same network model as the server on the microcontroller chip Cortex-M7 and load the derived weight value to realize the research of face recognition using neural network on the microcontroller.The main research work is as follows:(1)A method of building a neural network on a microcontroller chip for data processing and identification is proposed.The microcontroller chip not only has the advantages of low cost,low power consumption,long life,etc.,but also does not rely on the operating system,and is more stable and reliable,which can effectively solve problems such as excessive pressure on the background server.The application program interface of the CMSIS-NN neural network kernel can be easily directed to any neural network framework.The library contains optimized neural network functions to improve the recognition performance of neural networks on microcontroller chips.(2)Design and build a lightweight neural network model suitable for microcontrollers.A lightweight convolutional neural network was built aiming at the problems of the large scale of the existing network model,the large number of parameters,the complexity of the calculation,and the high requirements on the memory and power consumption of the device.The network has only five convolution layers and three pooling layers,which not only has a small amount of computation,but also has a low power consumption.(3)A sparse convolutional neural network is designed to tailor the model parameters.Because there is a lot of redundancy in the weight parameters of the trained neural network,a large number of parameters in each layer of the convolution kernel are useless.These useless parameters waste memory,increase the amount of calculation,and cannot express model features.Therefore,after the training of the convolutional neural network is completed on the server,the discard matrix is calculated by the neural correlation and sparseness of the models,and the discarded matrix is used to perform model compression on the trained neural network.Two kinds of databases were used for experimental tests.The recognition performance of the convolutional neural network built on the server,the neural network after model sparseness,and the sparse neural network built on the microcontroller were compared.The experimental data of the three are basically be consistent.Therefore,building a sparse neural network on the microcontroller chip for face recognition can effectively solve problems such as network bandwidth and excessive server pressure,which is of great significance for improving system security in various industries.