Face Image Inpainting Method Based On Generative Adversarial Network

Image inpainting is an important research hotspot in the field of image processing.The purpose is to use the existing structural information and internal information of the image to realize the self-repair of the image and achieve the ability of "self-recovery" or even "recreation".As the most widely used information carrier in daily life,face image involves the fields of face recognition,face detection,face editing and face prediction.However,whether in real life or in information transmission,the obtained face images are inevitably damaged or occluded,which brings challenges to face-related practical applications.Therefore,the research field of face image inpainting came into being.The purpose is to infer the content of the occluded area according to the existing information of the occluded face image,so as to realize the restoration of the image.This paper focuses on the basic topic of face image inpainting,and aims at three specific problems: 1.random occlusion of face image inpainting,2.mask-wearing face inpainting based on face guidance,3.multi-possible prediction of face image inpainting,and three novel solutions are proposed,and their main work and innovations are as follows:1.Aiming at the problem of face image inpainting with random occlusion,a thick and thin inpainting network is proposed,combined with the attention mechanism,to carry out two-stage image inpainting.Among them,the coarse inpainting network mainly focuses on the structural information and semantic information of the occluded image,and the fine inpainting network mainly focuses on the texture information of the occluded image.By adopting the U-Net network structure and combining the long and short attention mechanism,the context information of the image can be fully utilized in the two-stage inpainting process.By using gated convolutions instead of ordinary convolutions,self-learning of dynamic features is achieved,resulting in better inpainting results.Through experiments,combined with quantitative and qualitative analysis,the effectiveness and superiority of this method are proved.2.Aiming at the problem of mask-wearing face inpainting based on facial attribute guidance,this paper assumes that the basic features of the face under the mask are known,and the face attribute information is used as input to guide the restoration process of the network.Specifically,a two-way parallel neural network is proposed,including the reconstruction path and the generation path.At the same time,the attribute prior of the face is used as part of the input in the form of a vector to guide the direction generated by the network.Guided by the feature and distribution of the reconstructed path and the face attribute vector,the generated path fits the distribution,and then generates an image that conforms to the input face attribute.In order to simulate the scene of wearing a mask,this paper uses the Celeb A face dataset for pose estimation and positioning,and generates a oneto-one corresponding mask dataset.Through experiments,combined with the results of indicators and visualization,the effectiveness of this method is proved.3.Aiming at the multi-possible prediction problem of face image inpainting,two paths of diversity structure generation and diversity image generation are proposed.Among them,the diversity structure generation path adopts Transformer,and combines the bidirectional attention mechanism to capture the context information of the occluded area,so as to predict the probability distribution of the occluded area.The image generation path combines the probability distribution predicted by the diversity structure generation path to generate images.Through experiments,the diversity and complexity of the generated results are demonstrated by performing diversity index evaluation and visual evaluation on the generated images.