Robust Blind Deconvolution Using Relative Total Variation as a Regularization Penalty

Abstract

Blind image deblurring or deconvolution is a very hot topic towards which numerous methods have been put forward in the past decade, demonstrating successful deblurring results on one or another benchmark natural image dataset. However, most of existing algorithms are found not robust enough as dealing with images in specific scenarios, such as images with text, saturated area or face. In this paper, a robust blind deblurring approach is presented using relative total variation as a regularization penalty. To the best of our knowledge, none of previous studies have followed this modeling principle. The underlying idea is to pursue more accurate intermediate image for more reliable kernel estimation by harnessing relative total variation which could extract salient structures from textures and suppress ringing artifacts and noise as well. An iterative estimating algorithm is finally deduced for alternatively update of sharp image and blur kernel by operator splitting and augmented Lagrangian. Extensive experiments on a challenging synthetic dataset and real-world images as well are conducted to validate the new method with comparison against the state-of-the-art approaches from 2006 to 2014. Based on both PSNR and SSIM and a non-reference evaluation metric, the results demonstrate that our approach is comparatively more effective to process blurred images in more challenging scenarios.