부산대학교 도서관

It is known image priors are essential to blind deconvolution. Reweighted graph total variation (RGTV), as a new prior to substitute the most classic TV, is shown superior to TV as well as several other state-of-the-art models in terms of both theoretical and empirical performance. In this paper, we take a step forward providing a simpler geometric view to RGTV, instead of the previous graph spectral interpretation made in the graph frequency domain. In specific, we formulate blind deblurring just via use of a derivative of the Leclerc loss, which is geometrically proved an appropriate candidate to promote the piecewise smoothing and sharpening desired by RGTV. A by-product of such a perspective is to closely relate blind and non-blind deblurring in a fairly naive fashion. A fast algorithm is then deduced to update the sharp image and blur kernel alternately, through implementing our simplified RGTV as a reweighted-L1 regularizer rather than a graph L1-Laplacian regularizer. Numerous experiments on challenging blurred images show a much better performance of the proposed approach than original RGTV, in terms of both effectiveness and efficiency. Additionally, the proposed method achieves a comparable or superior performance to other state-of-the-art methods, either model-based or deep learning-based ones.