Research On Image Inpainting Via Multiscale Content Attention Equalization

Content-based attention image restoration technology refers to selecting similar pixel blocks from the background area to fill the pixels in the damaged area through the attention module of channel dimension,so as to realize the repair task.In recent years,a large number of excellent attention-based image restoration algorithms have been proposed.However,the repair results are prone to blur,artifacts and discontinuities between pixels.In the repair task of large-area damage,it is also easy to lead to unclear details of the generated image.The above factors hinder the research of the repair algorithm to further improve the performance.To solve the above problems,thesis analyzes the shortcomings of current image restoration algorithms,and gives the following research results:1)To solve the problem of discontinuous pixels and blurry artifacts in Deepfill algorithm,a feature balancing method is introduced in the content attention module to improve the inpainting performance of Deepfill.Feature equalization refers to the fusion of surrounding background features in the foreground pixels generated by attention,so that the generated foreground pixel points can get suggestions for the surrounding background blocks and enhance the continuity of foreground pixels.By comparing the repaired images visually,the experiment shows that adding a feature-balanced content attention module can generate images with clearer structure and fewer artifacts.In the numerical analysis,the improved model achieves an average SSIM value of 90.2% and an average PSNR value of 33.01 d B in the Celeb A-HQ dataset’s inpainting experiments for small area masks(<20%),which is the same as PEN-Net,the frontier repair algorithm.In tasks with larger area masks(40%-50%),the average PSNR of the module is 27.7d B,and the average SSIM is 86.74%,which surpasses the performance of several other frontier algorithms.The data show that the image quality generated by this algorithm is higher and closer to the original image.2)To further solve the problem of unclear details in image generation during irregular mask repair tasks,a multiscale image inpainting model using dilated convolution is proposed.On the basis of the improved algorithm,the dilated convolution is used to extract different scale features of the image,foreground attention features are computed at each scale branch,and channel dimensions are stitched together to provide suggestions at different scales for generating foreground pixels.The algorithm is compared with four other advanced classical models.The SSIM and PSNR scores of this algorithm are 26.1d B and 84.43% under irregular random masks with coverage over 40%,which surpasses those of other algorithms.At the same time,in the visual repair results comparison experiment,the repair results of this model are more natural,and less artifacts are produced under the large area of irregular mask.The experiment also validates the effect of the model on a wider range of repair tasks using PASCAL VOC2007,which covers 20 image types,and a large multi-scene Places2 dataset.The experimental results show that the improved module can effectively improve the repair quality and solve the problem of discontinuous pixels in the generated image,so as to reduce the problems of blur and artifact.The repair performance of this model is better than other algorithms in a large area,and surpasses other similar cutting-edge repair algorithms in various main data indicators.Finally,this paper also uses the improved algorithm to realize the system application for photo art,and realizes the functions of face refinement,photo target removal and so on.