A self-adaptive learning method for motion blur kernel estimation of the single image

Abstract

The estimation of blur kernel is the first and principal steps in the deconvolution of single blurred image. The quality of image restoration highly depends on its estimation accuracy. We then propose a new modified-Radon-transform approach along with a low-high-pass filtering method to estimate the motion blur parameters by a self-adaptive learning strategy, which greatly improved the deblurring quality of the blurred image. The Gaussian low-pass and high-pass filters are adopted to reduce the noise level in blurred image, and the batch normalization and self-adaptive method are considered to eliminate the interference from the noise stripes. It is noted that the estimation of blur angle plays an important cue for the exploration of blur kernel. The experimental evaluation is conducted on both synthetic VOC2012 database as well as the natural-real motion blurred single image with or without noise. The experimental results show that our proposed method can obtain more accurate and more reliable blur parameters than other approaches.