Research On Image Inpainting Algorithm Based On Generative Adversarial Network

This paper has carried out the research of image inpainting algorithm based on generative adversarial network(GAN).Aiming at the structure and texture restoration problems existing in current image inpainting tasks,high-quality image inpainting results can be obtained through the adversarial game training of generator and discriminator.This paper specifically carried out and completed the following research works:(1)Aiming at the problem of inconsistent structure and texture connection between the filled area and the known area in the inpainting result when the texture around the damaged area is complex,a two-stage image inpainting algorithm based on a gated convolution generative adversarial network is proposed.In edge repair network and texture repair network,gated convolution is used to replace the dilated convolution in the residual block of the network,which effectively learns the relationship between the background and the mask.Network uses the spectral normalized markov discriminator to accelerate the network convergence speed and stabilize the network training process.Numerical experimental results show that compared with the previous two-stage inpainting algorithm,when restoring images with different sizes of mask regions,the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)of the algorithm are respectively increased by 3.8% and 3%,and the subjective effect is significantly improved.(2)Aiming at the problem of how to better keep the structure consistency between the repaired area and the known area,a generative adversarial network image inpainting algorithm based on context normalization(CN)is proposed.Add a CN module after the residual block of the generator in the edge repair stage to solve the mean and variance drift problems;add a CN module after the residual block of the generator in the texture repair stage to correct the damaged area Perform separate normalization with the undamaged area,and perform a global affine transformation to improve the structure recovery ability;to ensure the stability of the predicted context area,the spectrum normalized Markov discriminator and the context discriminator are combined to perform optimization judgments.Numerical experiments show that evaluation indicators of the algorithm are better than the previous chapter.In terms of subjective visual effects,the algorithm improves the image structure restoration more obviously.(3)Carried out the engineering application demonstration design of the intelligent removal of image objects,and developed an auxiliary demonstration tool for intelligent image restoration based on the PyQt5 framework.The tool can run cross-platform and supports operating systems such as Windows,Linux or mac OS.Its main function has two: freely draw the mask area on the original image,complete and present the inpainting result in real time;or automatically recognize the mask area for inpainting and display the result.In addition,this tool also opens the interface for adding custom algorithms to meet the expansion and upgrade of subsequent algorithms.