Kernel Estimation and Deconvolution for Blind Image Super-Resolution

Jiali Gong, Hongfan Gao, Jiahao Chao, Zhou Zhou, Zhengfeng Yang, Zhenbing Zeng

Abstract: Blind super-resolution, different from conventional non-blind super-resolution based on the assumption of fixed degradation, handles various unknown Gaussian blur kernels, and thus is closer to real-world application. The accuracy of kernel estimation and deconvolution directly influences the performance of overall super-resolution results, but recent works usually introduce artifacts during the process. In this paper, we propose our methods of a more accurate kernel estimation module (KEM) and deconvolution module (DM). Additionally, KEM and DM are embedded in kernel estimation and deconvolution structure (KEDS), which improves the results to a large extent once combined with non-blind networks.