Research On Image Repair Algorithm Based On Generating Antagonistic Network

Image repair is a technique of filling in the missing areas of a damaged image with alternative content to make the image visually real and semantically correct,removing unwanted objects in the image or modifying damaged areas in the photo.Traditional image repair techniques usually use low-level features to match and paste similar pixel blocks to fill in the missing pixels,which is poor in the details of the area to be repaired and only suitable for the repair of small missing areas.In recent years,the deep learning method has achieved remarkable results in the image repair work,especially the proposed generated antagonistic network makes the image repair technology has achieved exciting results,therefore,this paper proposes a deep learning image repair algorithm based on the generated antagonistic network.Firstly,through the in-depth study of the literature and the research on the theoretical basis of deep learning of image repair,the framework and working principle of the convolutional neural network are comprehensively studied,and the method of using the expanded convolution in the convolutional layer and the working principle of generating the antagonistic network are expounded.In order to make the local detail generation more reasonable and make the global feature and local texture of the image approximate to the real repair effect,an improved image repair algorithm is proposed.That is,local discriminant network is introduced on the basis of generating antagonistic network,so that the network model includes generative network,global discriminant network and local discriminant network.The convolutional layer of local discriminant network is improved,and the expanded convolutional layer is used to replace part of the convolutional layer to extract the local image information of a larger area.Then,the perceptive loss function is introduced,and the difference of feature graph of VGG19 network activation layer based on pre-training is compared.The mean square error loss function,perceived loss function and countermeasure loss function are used in the network to generate countermeasure network.The experimental results show that the proposed method is compared with the typical Patch Match and Context Encoder algorithms in the open celeb A data set,and the proposed algorithm has some improvement in image repair effect from the subjective and objective evaluation indexes.