Design And Implementation Of Lightweight Face Recognition Neural Network System Based On MobileNet V2

Security has always been the focus of attention because it related to the safety of people’s lives and property.Face recognition in the field of security is also the focus of computer vision research because it is the cornerstone of many security scene applications.However,in recent years,with the development of neural networks,the network structure has been continuously widened and deepened.Although the trained model greatly improves the recognition accuracy,it will consume a lot of memory and computing resources,which makes it difficult to deploy in hardware resources that are limited Mobile.Therefore,designing a lightweight neural network and obtaining a high-performance,low-consumption model is the key to solve the problem of applying face recognition to security of mobile end fields.MobileNet is a lightweight deep neural network proposed by Google for mobile devices.Application scenarios include object detection,fine-grained classification,face attributes,and large-scale geographic position.MobileNet is a deep network,but compared with other deep neural networks,it can significantly reduce model parameters,improve computing efficiency,and train high-performance,low-consumption models.However,the MobileNet network has not reached the optimal level,and the input and output layers of the network will still generate a large number of calculations in this application.The MobileNet network is improved to lightweight neural network in this thesis,it is applied to security in the field of face gender and age classification research and can solve the problem that face recognition is difficult to apply to mobile terminals.The main work in this thesis is as follows:(1)The popular MobileNet v2 network is improved and integrated into the s-MobileNet v2 network,which is used to the study of face gender and age classification in the face image data set in the security field.(2)Firstly,the MobileNet v2 network is improved in this thesis to solve the problem that a large number of convolution calculations in the MobileNet v2 network in the input layer and the output layer.The h-Relu6 function is introduced in the input stage to improve network efficiency,reduce channels and network parameters of the input layer;the average pooling operation is advanced in the output stage,which greatly reduces the calculation amount of the output layer.Secondly,in order to improve network performance,a SE module is introduced on the bottleneck block with a step size of 1 to automatically optimize the deep network parameters.Finally,in order to enhance the feature extraction effect,an information aggregation module is added after the output layer of the bottleneck block to perform feature extraction on the output image in multiple dimensions.(3)For different classification tasks of face gender and age,combining with the characters,different dataset is constructed.Besides this,data augmentation technologies to expand the data set and traditional image processing techniques(randomly changing the contrast,saturation,brightness,chroma,etc.)are used to solve the problem of a single image acquisition path to improve the robustness of the model,different evaluation functions are designed in test sets to evaluate the experimental results of different tasks.(4)Experiments on the s-MobileNet v2 network proposed are performed in this thesis.Compared with the MobileNet v2 network,the memory of the gender recognition model finally trained in this thesis is reduced by 48.8%,the computing performance is improved by 17%,accuracy is increased by 3%;the average error of age classification recognition decreased from 5.19 to 3.1.The effect of s-MobileNet v2 network is better than that of MobileNet v2 network,and it can better solve the problem of face recognition applied to mobile terminals.