Abstract
Motion blur is visible whenever the shutter speed of a camera is slow compared to the speed of unintended camera motion. General motion blur is a very complex type of blurring, and state-of-the-art blind image deconvolution methods rarely produce adequate results due to the ill-posed nature of the problem. Even modern deep-learning algorithms sometimes fail at the task. Modern deblurring approaches typically use a series of noisy images with shorter exposure time for the reconstruction of a high quality image. However, even with a shorter exposure time some blurring still remains. The good news is that, with little time for the motion vector to change direction significantly, this particular type of motion blur is much easier to model. The crucial stage in any deblurring process is the estimation of blur parameters. In this article we present a patch-based linear approximation to motion blur with the focus on effective estimation of the direction of linear blur. We use a CNN model for estimating the parameters of a linear blur kernel for each 32 × 32-pixel patch of an image and calculating a confidence value for each patch.

This publication has 11 references indexed in Scilit: