E2-RealSR: efficient and effective real-world super-resolution network based on partial degradation modulation

The goal of efficient and effective real-world image super-resolution (Real-ISR) is to recover the high-resolution image from the given low-resolution image with unknown degradation by limited computation resources. Prior research has attempted to design a fully degradation-adaptive network, where the entire backbone is a nonlinear combination of several sub-networks which can handle different degradation subspaces. However, these methods heavily rely on expensive dynamic convolution operations and are inefficient in super-resolving images of different degradation levels. To address this issue, we propose an efficient and effective real-world image super-resolution network (E2-RealSR) based on partial degradation modulation, which is consisted of a small regression and a lightweight super-resolution network. The former accurately predicts the individual degradation parameters of input images, while the latter only modulates its partial parameters based on the degradation information. The extensive experiments validate that our proposed method is capable of recovering the rich details in real-world images with varying degradation levels. Moreover, our approach also has an advantage in terms of efficiency, compared to state-of-the-art methods. Our method shows improved performance while only using 20 link as open source.