야간 영상에서의 비균일 블러 제거

Abstract

This thesis presents an algorithm to remove non-uniform motion blurs of low-light images. Low-light images have less edge information than natural images, so MAP (Maximum A Posteriori) approaches using salient edge information often fail to estimate blur kernels in low-light images. We use light streaks in low-light images as additional information which shows camera motion during exposure.We extract light streaks using several properties and initialize blur kernel with extracted light streaks. Then, we alternatingly restore blur kernel and latent image. In the blur kernel estimation step, we estimate a blur kernel using both edge information and light streaks. In the latent image estimation step, we deconvolve the blurry image using the estimated kernel with outlier handling. After several iterations, we obtain the final blur kernel and deconvolve the blurry image using the final blur kernel. Experimental results demonstrate that our method obtains good results on both synthetic and real world examples.