Extraction of Degradation Parameters for Transparency of an Image Restoration Network
Kazutaka Uchida, Masayuki Tanaka, Masatoshi Okutomi · 2019
Many image restoration processors based on convolutional neural network (CNN) has been proposed because of its high performance. However, it is well known that restoration by these networks is not robust against perturbations on a degradation model. If restoration fails, it is difficult for users to find the cause because no explanation is given by the network. In this paper, we propose an additional network to extract internal parameters for an image restoration network to supply users explaining information on restoration process. Experimental results show that the proposed network successfully extracts estimated degradation attributes and gives helpful information to assist users to find a root cause of a restoration failure.