Research On Face Recognition Based On CNN And Its Embedded Application

Face recognition,as a highly practical and reliable biometric recognition technology,is widely used in daily life.With the rise of Convolutional Neural Networks(CNN)nowadays,high-precision face recognition models continue to appear,and face recognition has ushered in a new round of development peaks.However,face recognition technology based on convolutional neural networks faces an important constraint because the model generated by the network is large and the calculation is complicated.Although this type of face recognition technology can run in real-time on a highly-configured computer,it is difficult to apply it to an embedded platform without a graphics processing unit(GPU).Aiming at the above problems,this paper studies from two aspects of face detection and face recognition.For face detection,a face detection algorithm based on Lw-YOLO(LightweightYOLO)is designed.This algorithm uses depthwise separable convolution in the basic CNN to replace the traditional convolution method,which effectively reduces the amount of network parameters.At the same time,multiple 1*1 point convolutions are used to increase the number of channels to obtain more features.By improving the multi-scale prediction method,the prediction of three scales is reduced to two scales,and the recognition accuracy of the network is ensured with a reduced amount of parameters.The experimental results show that the size of the trained Lw-YOLO network model is only 3.1M,and an average accuracy of 77.13% is obtained on the FDDB dataset,and the detection speed is faster than that of similar networks.In face recognition,a face recognition algorithm based on S-GoogLeNetv3(Small-GoogLeNet version3)network is proposed.This paper improves the GoogLeNet network by replacing large convolution kernels with small convolution kernels and trimming auxiliary classifiers to reduce the amount of parameters.At the same time,ArcFace which has better classification in angular space,is selected as the loss function.The face recognition model is obtained by training on the CASIA-WebFace face dataset,and the performance was evaluated on the LFW testset.It is found through experiments that the accuracy and time required for the model achieves the intended design purpose.Finally,this paper uses the GPU-less Raspberry Pi 3B+ development board as the embedded platform,uses PyQt4 to design the interface,and combines Darknet,Torch and other deep learning frameworks to design and implement an embedded face recognition system based on CNN.The system uses lightweight CNN face recognition technology,which reduces the amount of network parameters and accelerates the speed of face recognition while ensuring recognition accuracy.The use of convenient embedded devices for face detection,self-built face dataset training,face recognition and other integrated operations is in line with the trend of embedded face recognition today.