Blind Super Resolution with Feature-Oriented Adaptive Degradation Adjustment

Conventional methods in blind super-resolution first estimate the unknown degradation kernel from the low-resolution image and then leverage the degradation kernel for image reconstruction. Such sequential methods always have two basic weaknesses. Firstly, the lack of robustness which is due to the severe performance drop when the estimated degradation kernel is inaccurate. Another is the failure to effectively utilize the degradation kernel to achieve super-resolution reconstruction due to the domain gap between the degradation kernel and the image feature. To address these issues, we propose a blind super-resolution framework with Feature-Oriented Adaptive degradation adjustment network (FOAnet). Specifically, We design a novel kernel estimation network using U-Net style, which can greatly improve the performance and accuracy of kernel estimation with the powerful extraction capability by fusing features from different levels and channels. In addition, we start from the perspective of what kind of degradation is needed for current image features, adaptively adjust the predicted degradation kernel according to image features, and generate local dynamic filters and channel coefficients to modulate image features in order to flexibly handle the domain gap between degradation kernel and image feature. Numerous experimental results on synthetic data including Gaussian8, DIV2KRK, and real scenes demonstrate that the proposed FOAnet achieves state-of-the-art performance.