Multi-DIP: A General Framework for Unsupervised Multi-degraded Image Restoration

Most existing image restoration algorithms only perform a single task. But in the real world, the degradation pattern could be much more complex, such as blurred images that have been smudged or images with haze that have been blurred, and we call it multi-degradation. Many of these degenerations are coupled with each other, making it impossible to restore images by merely stacking the algorithms. In this paper, we propose Multi-DIP that uses DIP networks to solve the multi-degradation problem. We integrate multiple image restoration tasks into a unified framework. However, multi-degradation can cause difficulties for DIP networks to extract image priors. To alleviate this problem, we design a multi-scale structure to stabilize and improve the quality of generated images. We implement two image restoration tasks with the proposed DIP framework: deblur + inpainting and dehaze + deblur. Extensive experiments show that our proposed method achieves promising results for restoring multi-degraded images.