Unsupervised Image Restoration Via Deep Image Prior

The image restoration algorithm is a low-level computer vision task that can intuitively improve images’ visual results.In addition,the restored images can provide more reliable input for downstream tasks,which has broad application value and is a critical optimization node in computer vision engineering practice.The supervised image restoration algorithm can lead to significant performance degradation when the training set does not match the image domain of test samples,so the unsupervised approach has developed rapidly in recent years.This paper summarizes the implementation of an unsupervised image restoration framework based on the deep image prior for the single-mode degradation restoration task and improves the dehazing algorithm under this framework.This paper proposes a multiscale initialization method to limit the parameter solution space to address the difficulty of training instability due to too many generators in the network.It uses the color attenuation prior to constrain the training process of the airlight map in the small-scale prior extraction stage.This paper compares the method with other classical unsupervised image dehazing algorithms on the HSTS test-set and achieves better objective evaluation results and subjective visual results.In addition,to address the problem of more complex degradation patterns of images in the real world,this paper proposes an unsupervised image restoration framework adapted to multi-mode degradation,MultiDIP.It represents the image prior used to be obtained by excavating external datasets by a generator network structure and directly models the degradation process of images,using a modular manner to decouple multimode degradation.In this paper,we propose two specific implementations of the multi-mode degradation image restoration tasks "deblurring+inpainting" and "dehazing+deblurring" in the Multi-DIP framework and use a multiscale initialization method similar to that of the single-mode degradation restoration tasks.Due to the lack of publicly available datasets,the test sets matching the tasks are designed,and qualitative and quantitative experimental analyses are conducted on these two tasks to demonstrate the effectiveness of Multi-DIP.