Research On Face Recognition Algorithm In Multi-position

As the application of face recognition technology in our daily life becomes more and more popular,face image processing technology based on specific usage scenarios has emerged.In the specific face recognition process,the difficulty of face recognition caused by posture changes is a problem that needs to be solved urgently.When the human face poses a large angular deflection,the characteristic information of the face will be lost due to the angular deflection,which will affect the face recognition process.Therefore,it is an important research direction to realize high-precision face recognition under various pose changes through various methods.In order to solve this problem,a face frontalization algorithm can be used to correct profile faces,and the corrected positive picture is used as input for face recognition.In addition,the problem of face recognition due to image resolution is also serious,and restoration of high-quality face images has also become a research hotspot.Face correction algorithms mainly include traditional mapping methods,3D modeling-based correction methods,and deep learning methods.This article first introduces the Basic process of face recognition,the background technology knowledge of and deep learning,and then introduces the face frontalization algorithms based on deep learning,and explains the face correction algorithms SPAE,TP-GAN,and DR-GAN based on deep learning methods.An adversarial generative face frontalization network based on pose residuals FCM is proposed.The network is based on an adversarial generative network and includes a pose residual generator and a double discriminator(attitude condition discriminator,identity condition discriminator).The pose residual information and the prior face identity and pose information can guide the generation of frontal faces and improve the fit of the generator to the identity and pose.And designed a pose normalization module FTM to generate two-dimensional keypoint information of the face.Adding dense connections to the decoder structure of the generator helps reduce training parameters,avoids disappearance of gradients,and has a regularization effect,which effectively avoids overfitting.For low-resolution face images,a high-quality face generation network HFG is proposed.The model's training process adds the face correction feature map mentionedabove as a guide for training a Generating network structure for both face correction and face restoration.Finally,in the experimental stage,the CMU Multi-Pie data set was used to train the face correction model,and tested on the CMU Multi-Pie and LFW data sets.It was compared with today's excellent face correction algorithms to analyze the effect of the generated image;And in order to analyze the role of each structure in the model,comparative tests were carried out separately.HFG was trained using the CASIA-Webface dataset and tested on the LFW and IJB-C datasets.The proposed high-quality face generation models HFG and HFCM are compared with the best existing face restoration methods,and they are tested on the recognition network Resnet-50,and the face recognition rate after repair is analyzed.