Research On Key Techniques Of Image Inpainting Based On Domain Features

Image inpainting is one of the most popular research directions in the field of computer vision,and its application fields are very extensive.At present,deep learning method based on convolutional neural network is widely used in the field of image inpainting,which greatly improves the effect of image inpainting,but it still has the following problems:the content generation results of the image missing area are uncontrollable,the generated image's quality has artificial traces,the whole algorithm is not robust enough,etc.The thesis studies the above problems,and the specific work is as follows:1)Image inpainting algorithm based on perceptual domain-aware features prior is proposed.This algorithm uses the generative adversarial network and conditional network as the perceptual prior to constrain the content generation network of the missing area,and through content constraintion and texture constraintion improves the image inpainting performance.Compared with the traditional image inpainting algorithm,the image-aware inpainting algorithm of the domain's perceptual priori features has obvious improvement in the inpainting performance,especially in the image inpainting of missing key information.2)Unsupervised image inpainting algorithm based on domain features is proposed.This algorithm repaired the image through the auto-encoder network.We adopted missing content constraint,the adversarial constraint,the background reconstruction constraint and the smooth constraint in repairing process.The performance of a variety of a priori constraints greatly improve the quality of image inpainting.The algorithm does not need to give a clear standard answer,and can solve the ill-posed problems.It can produce good image inpainting performance in multiple datasets and its strong robustness.