Double Pyramid Field-aware Blind Deconvolution Framework

Blind deconvolution is a challenging problem of recovering a signal from their noisy convolution and it is prevalent in various fields including astronomical imaging, medical imaging, signal processing and computational optics. This paper discusses a more practical blind deconvolution problem which copes with the noisy convolution by spatially variant kernels. We found that the simple scheme used to participate the blurred images into regions for estimating spatially variant kernels is inaccurate. Otherwise, without using efficient kernel refinement and boundary condition, most of the current algorithms suffer more artifacts. In this paper, we proposed a double pyramid field-aware blind deconvolution framework. The proposed framework divides the given image to several patches by different field and estimate the coarse kernels of those patches by total variation-based algorithm. Then a new kernel refinement algorithm is proposed to refine the coarse kernels. After that, the non-blind deconvolution method with Hyper-Laplacian priors is used to get the deconvoluted whole image by their own refined kernels. Finally, the high-quality image is reconstructed by interpolating the several patch-kernel restored images together with the field-weighted interpolation method. Experiments illustrated that this framework can alleviate the restoration inaccuracy by single kernel, handle the unknown distortions to kernels and significantly improve visual quality.