Research And System Design Of Lightweight And Efficient Face Recognition Algorithm Based On Deep Learning

With the demand of CPU devices for deep learning algorithms,the amount of network model parameters and operating speed have become the difficulties that restrict the deployment of deep learning algorithms on ordinary devices.Existing networks are mostly bloated and cannot meet the usage conditions of CPU devices.Therefore,how to make the deep learning network more lightweight and efficient has become a hot field in the field of deep learning.The main research focus of this paper is how to design a basic network with few parameters and high accuracy to realize real-time high-precision face recognition without GPU devices.The main content of this article is as follows: In order to reduce the size of the model and improve the speed of the network,the original MobileNet V2 network structure was simplified to make it lightweight,efficient,and low memory access consumption.First,the unnecessary residual blocks in the original network are discarded to reduce the number of convolutional layers of the network and the amount of network parameters.Then by reducing the expansion speed of the original network convolution channel,to reduce the network's memory access,and improve the network's operating speed.In order to further increase the operating speed of the network,reduce the expansion coefficient in the residual structure,and modify the expansion method when the channel is expanded,so that it follows the parallel expansion method,so that the network has a smaller memory access cost and increase the actual operation speed.Finally,in order to make the accuracy of the network high,the method of fusion of spatially separable convolution and depth separable convolution features is adopted,so that the two can compensate each other in features.In addition,Arcface loss is used to replace the original SoftMax loss,and the features extracted by the network are more separable and robust by increasing the ability to constrain the network.By adopting the above strategy,the improved LE-Net network based on MobileNet V2 has achieved the same training conditions.The network model size is reduced to 2.3MB,the LFW test accuracy is99.53%,and the model speed is five times that of MobileNet V2.Finally,based on LE-Net,a real-time lightweight and efficient face recognition system isdesigned.The system includes a face detection part,face cropping and standardization processing,face feature extraction and comparison,face registration and deletion,System settings and other functional modules.Face detection uses MTCNN five-point face detection algorithm.The recognition algorithm uses LE-Net as the feature extraction network and uses the cosine metric as the feature comparison method of the system.On the Intel Core i7-4790 CPU,the average running speed of the system tested on some Megaface and CFP-FP data sets is 75 ms,and the accuracy is 99.20% and 89.27%,respectively.It achieves low memory consumption in real time.Face recognition.The experimental results of this algorithm on common CPU devices verify the effectiveness of this algorithm..