Kernel-aware Burst Blind Super-Resolution

Burst super-resolution (SR) technique provides a possibility of restoringrich details from low-quality images. However, since real world low-resolution(LR) images in practical applications have multiple complicated and unknowndegradations, existing non-blind (e.g., bicubic) designed networks usuallysuffer severe performance drop in recovering high-resolution (HR) images. Inthis paper, we address the problem of reconstructing HR images from raw burstsequences acquired from a modern handheld device. The central idea is akernel-guided strategy which can solve the burst SR problem with two steps:kernel estimation and HR image restoration. The former estimates burst kernelsfrom raw inputs, while the latter predicts the super-resolved image based onthe estimated kernels. Furthermore, we introduce a pyramid kernel-awaredeformable alignment module which can effectively align the raw images withconsideration of the blurry priors. Extensive experiments on synthetic andreal-world datasets demonstrate that the proposed method can perform favorablestate-of-the-art performance in the burst SR problem. Our codes are availableat \url{https://github.com/shermanlian/KBNet}.