Sharp image estimation from a depth-involved motion-blurred image

When capturing image of a 3D scene with depth variation, the image blur due to camera shake is no longer spatially invariant but spatially varying which contains the information of the depth of the scene. In this paper, we consider the task of latent image restoration from a single depth-involved motion-blurred (DIMB) image. To easily describe the depth effect, we propose a new framework to model the image blur where the blur kernel is parameterized by the motion curve of the camera and depth value of the scene. Moreover we present an iterative algorithm to alternatively estimate the depth map, motion curve and latent image, based on the minimization of a cost function. To make the estimation of depth map more accurate, a two-stage depth estimation algorithm is also proposed. In the experiment part, the performance of the proposed method is evaluated on many examples including synthetic and real-world cases and shows that our method can produce better results on the DIMB images. (C) 2015 Elsevier B.V. All rights reserved.