Research On Face Image Inpainting Based On Deep Learning

Face is an important way for human to convey messages and the most direct immediate representative of identity.By utilizing the features like specificity and traceability of face images,it is possible to identify the relationship between face and identity through the face recognition technology.face recognition is currently the most convenient and effective approach to identity authentication.It is widely applied in ticket checking in railway stations,unlocking mobile phones and checking on work attendance.With the COVID-19 spread of the epidemic in recent years,wearing a mask has become the main epidemic prevention measure,and the large area of covering face has reduced greatly the accuracy of face recognition.In order to reduce the chance of secondary transmission and infection when people remove their masks for face recognition in a public environment during the epidemic,it is necessary to improve further the technology and accuracy of face recognition.The main work of this paper is as follows:In the view of the current problem of low accuracy rate of face recognition system to identify people wearing masks,the face recognition rate is improved by building a more targeted face image inpainting system before the face recognition system to inpainting face image with high precision and inpainting as much as possible.The pursuit of high precision and face image inpainting has been one of the hot topics for many researchers in recent years,but it is difficult to improve the authenticity,controllability and clarity of face image because of the complex diversity of faces.For the above problems and challenges,this paper combines Multi-task Convolutional Neural Network(MTCNN)and Generative Adversarial Networks(GAN)technologies on the basis of deep learning and proposes an efficient face image inpainting model from two aspects of face recognition and localization and face image inpainting.The accuracy of face image inpainting position depends on the accuracy of face recognition positioning.For the problems of false face detection and inaccurate face features in the process of face image inpainting,a face recognition and localization model based on Multi-task Convolutional Neural Network was proposed.Firstly,introducing image classification Inception V2 module and adding batch normalization operation to the structure to extract more feature information while deepening the dimensionality of feature mapping to improve its accuracy of five features location;secondly,we can choose Mean Square Error(MSE)loss function and optimize its correlation coefficient to improve the learning efficiency and stability,at the same time,we incorporate recognition formula of face false detection to preserve the correct face information.For the problems of image distortion,blurring and low accuracy of five features inpainting in the process of face image inpainting,a face image inpainting model based on Generative Adversarial Networks was proposed.Firstly,the structure and parameters of the generative and discriminative models are designed to increase the stability of the structure,introduce full connection layer per channel and normalization layer to reduce the probability of losing image features and improve the convergence speed of model training.Second,tanh function and Leaky Re LU function are introduced into the activation function,and an improved Sigmoid function is proposed to expand its unsaturated region,which solves the problem that the generation countermeasure network is prone to gradient disappearance or gradient explosion.Meanwhile,the adaptive gradient optimization algorithm is used to replace the commonly used stochastic gradient descent algorithm to improve the image fidelity;Wasserstein distance is introduced on the basis of self adversarial loss function,the total variation loss function and the reconstruction loss function are introduced to improve the clarity of the inpainting image and solve the problem of image distortion.After the above analysis,the paper carries out experiments and makes contracts for several face image inpainting algorithms.Results show that the optimized face image inpainting model could lead to solve the problem that generative adversarial network training is too free,more accurate inpainting of five features and further improved forecast effect,it is more conducive to face identification.