Research On Face Repair Algorithm Based On Deep Generative Adversarial Model

In real life,the breakage of face images has caused major problems in all walks of life,and face image repair technology can solve this problem intelligently and efficiently.Face image repair is a computer vision technique that uses mathematical model algorithms to model the remaining available image regions to estimate the pixel distribution of the face in the defect area.Face image repair technology can save a lot of time and manpower for all walks of life,so the research of this technology has important practical significance.Face image repair technology can be widely used in the fields of archaeology,security criminal investigation and face recognition.At present,deep learning has achieved important research results in the field of computer vision.The mainstream deep generative adversarial models have their own focuses when applied to face repair,and they can not repair the damaged face of large random area.In view of this problem,based on the in-depth study of the powerful learning ability of the deep generative adversarial models,this paper proposes a new face repair algorithm Face R-Net based on the existing excellent research results.The main contents of the algorithm are as follows:(1)The Face R-Net network model proposed in this paper first preprocesses the image to be repaired,and fills the damaged area with the remaining pixels outside the damaged area,so that the subsequent model training can make full use of the priori pixels of the known area in the damaged image,learning high-level semantic information at the pixel level,achieving the goal of making model training more efficient and better.The experimental comparison shows that pre-processed Face R-Net has better repair effect on face-damaged images than the algorithm based on deep generative adversarial model that uses white noise to pre-process the damaged area of the face.(2)The face repair model based on Deep Convolution Generative Adversarial Networks(DCGAN)uses a single discriminator network to discriminate the authenticity of the generated face image,which makes the generator network not get enough adversarial information,so that training slow,poor training results.The Face R-Net algorithm proposed in this paper improves the deep convolution generative adversarial networks,by adding the global discriminator network and the local discriminator network to learn the features of the entire face image and the local defect area,the global discriminant loss and local discriminant loss are obtained,and the mean square error loss is suppressed to avoid network learning from being too conservative,and the edge loss is added to make the model more consistent in repairing the broken edges of the face.Compared with several mainstream face repair algorithms based on deep generative adversarial model,the proposed method has better effect and wider application range.(3)In order to solve the problem of unnatural boundary transition in the process of face image repair,this paper proposes two kinds of face image post-processing algorithms based on poisson fusion and fast marching method,both of which make full use of the priori information of the face damage image and fuse with the face repair image,so that the resulting face repair image is more reasonable and more realistic.The comparative experimental analysis shows that both post-processing algorithms can get better face-repair image fusion effects.