Self-Supervised Cross-Scale Nonlocal Attention Network for Blind Image Super-Resolution

Most existing image super-resolution (SR) meth-ods commonly assume that the degradation kernel is fixed and known. Blind SR aims to handle various unknown degradation processes closer to real-world applications and more generalizations. We propose a self-supervised cross-scale nonlocal attention network for blind SR (CNSR) which jointly models a blur kernel estimation module (KEM) based on a regularization model and a high-resolution image reconstruction module (HRM) based on a deep neural network. The low-resolution (LR) image is used as the supervision signal, and the blur kernel and high-resolution image are estimated simultaneously by iterating the two modules alternately. In HRM, we introduce a cross-scale nonlocal correspondence aggregation module (CNCAM) that uses the cross-scale self-similarity of images to provide additional information for image reconstruction. Experimental results show that CNSR can effectively improve image reconstruction performance.