Face Feature Extraction Based On Depth Convolution Neural Network

Face recognition technology is one of the most challenging research fields in image processing,pattern recognition and artificial intelligence,and its practical application has a bright future.A typical face recognition system consists of the following six parts: preprocessing,face detection,biometric detection,facial feature point location,feature extraction,matching and recognition.Feature extraction has the most direct impact on the accuracy of face recognition.With the birth of deep convolutional neural networks,classic networks such as Le Net,Alex Net,VGG,Google Net and Deep ID2 appear with a recognition rate of more than 99%.After the birth of chip-level deep learning platforms such as Caffe2,Caffe2 go and even FPGAs,Deep Convolutional Neural Networks Mobile Net,Shuffle Net Appear.In this paper,the method of face feature extraction based on the deep convolution neural network,the main research of it is the following aspects:(1)Had studied the current mainstream DCNN structure,especially Mobile Net and Shuffle Net at light-weight mobile level,and found some efficient and practical structural features of convolution.A face feature extraction method based on deep convolution neural network is proposed,and a non-linearly separable deep convolution neural network X-Net is designed.X-Net belongs to a lightweight convolutional neural network,which increases the spatial correlation of depth convolution to a certain degree.The detail features and the spatial correlation features are well preserved because conv3*3/1*1 Structure characteristics,the number of facial feature extraction network calculation and the number of parameters have also been better controlled.(2)Aiming at the imbalance of sample in the actual face data set,Focus loss is introduced to reduce the effect of sample deviation on network fitting.In addition,this paper also summarized and analyzed the current mainstream live detection technology,and proposes a Caffe-based network encryption method to ensure the security of facial feature extraction.In-depth study of deconvolution visualization and a series of training methods and some pre-processing techniques,facial feature extraction more engineering optimization.