Un-Paired Real World Super-Resolution with Degradation Consistency

Abstract

With the development of deep convolutional neural networks, deep-learning based image super-resolution has achieved excellent performances with paired training samples and prior degradation models. Nevertheless, in real world, the information transportation and compression procedure are generally unknown and only the low-quality low-resolution images are available. In this case, un-paired real world image super-resolution task is far more challenging than the paired one. In this paper, we develop an efficient un-paired super-resolution method with degradation consistency (DCSR). Specifically, a multi-level aggregation network (MLAN) is developed for feature representation, and three degradation-consistency losses are introduced for synchronously retaining the inherit contents and generating desired photo-realistic details. The proposed methods show superior performance on benchmark datasets and achieve 2nd place on "Target Domain RWSR" track of the AIM RWSR Challenge [19].