Space-variant blur kernel estimation and image deblurring through kernel clustering

This paper presents a space-variant blur kernel estimation and image deblurring framework. For space-variant blur kernel estimation, the input image is divided into small patches, and for each patch, the blur kernel is estimated. The estimated kernels are then grouped to determine different kernel clusters in the image. During clustering, unreliable kernel estimates are eliminated. The blur kernel for each kernel cluster is finally refined using the corresponding image region, which is the union of image patches associated with the kernels in the cluster. For space-variant image deblurring, the entire image is deconvolved with each blur kernel to produce a set of deblurred images. These images are then fused to produce a blur-free image, where the fusion process selects the optimal regions from the set of deblurred images.