Generative adversarial network image inpaiting based on diffusion process
To address the issues of blurry texture of the repaired images and instable training process in existing image inpainting algorithms,this paper proposes a Generative Adversarial Network(GAN)based image inpainting approach leveraging the diffusion process.By incorporating the diffusion model into a dual-discriminator GAN,the generated images from the generator and real images undergo a forward diffusion process to obtain the inverted ima-ges and real images with Gaussian noise.These images are then fed into the discriminator to enhance the inpainting quality and improve the model training stability.Style loss and perceptual loss are introduced into the loss function to learn semantic feature differences,eliminate motion blur,and preserve more details and edge information in the in-painting results.Qualitative and quantitative analyses,along with ablation experiments,have been conducted on the datasets of Celeb A and Places2.The evaluation and restoration outcomes show the superior performance of the pro-posed approach.Compared with current inpainting methods,the proposed approach achieves an average improvement of 1.26 dB in Peak Signal-To-Noise Ratio(PSNR)and 1.84％in Structural Similarity Index Measure(SSIM),while reducing the L1 error by an average of 25.7％.Furthermore,the changes in the loss function indicate that the image inpainting algorithm with diffusion process exhibits more stable training behavior.
万方数据
维普
