Progressive Image Restoration with Multi-stage Optimization

With the recent development of deep learning technology, researchers have achieved significant results in small-scale image inpainting. However, when the missing area is large, undesirable artifacts and noise are introduced into the inpainting area. Hence, we present a multi-stage progressive image inpainting framework based on the well-known generative adversarial network(GAN) to solve this problem. In our MOPR-GAN method, generator uses a progressive inpainting module(PIM) and an image optimization module(IOM), while discriminator combines a patchGAN with an attention mechanism and a globalGAN. The PIM can gradually repair the image loss area and generate an attention map simultaneously. The IOM optimizes the details of the generated image based on the information provided by the attention map. The discriminator can capture the local continuity and universal global features of the image better. When comparing the test results with the latest research, the model showed a significant effect in both qualitative and quantitative analyses.