Key points are not available for this paper at this time.
Recent blind super-resolution (SR) methods typically consist of two branches, for degradation prediction and the other for conditional restoration. , our experiments show that a one-branch network can achieve comparable to the two-branch scheme. Then we wonder: how can one-branch automatically learn to distinguish degradations? To find the answer, propose a new diagnostic tool -- Filter Attribution method based on Integral (FAIG). Unlike previous integral gradient methods, our FAIG aims at the most discriminative filters instead of input pixels/features for removal in blind SR networks. With the discovered filters, we develop a simple yet effective method to predict the degradation of an image. Based on FAIG, we show that, in one-branch blind SR networks, 1) are able to find a very small number of (1%) discriminative filters for each degradation; 2) The weights, locations and connections of the filters are all important to determine the specific network. 3) The task of degradation prediction can be implicitly realized by discriminative filters without explicit supervised learning. Our findings not only help us better understand network behaviors inside one-branch SR networks, but also provide guidance on designing more efficient and diagnosing networks for blind SR.
Xie et al. (Mon,) studied this question.