| The traditional digital image inpainting technology based on structure and texture only reconstructs the damaged image at the pixel level and lacks the understanding of the high-level semantics of the image.Therefor,it is unable to repair the image with a large missing regions.Image inpainting technology based on deep learning shows a great potential in restoring large missing regions by using the strong semantic elaboration and feature extraction ability of generative adversarial network.However,most existing works either merely consider the information outside the missing regions or rely on regions context to generate semantically coherent patches while ignoring the semantic relevance and feature continuity inside missing regions,which may lead to a discontinuous contextual structure and blurry texture details.Aiming at the above shortcomings,this paper has carried out in-depth research on image inpainting based on deep learning,the contents are as follows:(1)Based on the generative adversarial network,a cascade inpainting framework from coarse to fine is introduced.The coarse inpainting network first performs a preliminary prediction and restoration of the image to stabilize the training process and expand the receptive field,and at the same time provides the prior distribution of the image structure for the refine network;the refine network can learn the structure information of the image more comprehensively according to the results of the coarse network so as to improve the performance of whole model.(2)To solve the problem of semantic distortion in image reconstruction,an attention model of internal and external context is proposed,which is embedded in the fine repair network.This model can not only maintain the high correlation between the internal and external pixels in the missing area,but also enhance the semantic continuity between the generated pixels in the missing area,and combine the pixels inside and outside the missing area to guide together Reconstruction of the missing area.(3)The adversarial loss based on natural image level is not enough to produce clear texture details.In order to solve this problem,this paper introduces a texture adversarial loss based on wavelet decomposition to enhance the details of the image.Through wavelet decomposition of the generated image,the high-frequency components representing the image details are obtained,and then the high-frequency components are used as the input of the discriminator to learn the texture details of the image.In this paper,two large image datasets,Celeb A and Place2,are used for experiments.The experimental results show that the algorithm can reconstruct the results consistent with visual reality no matter for rectangular or irregular missing areas,and compared with some existing methods,the results of the algorithm in this paper have more continuous semantics and clearer texture details. |
