A Fast Approach for Video Deblurring using Multi-Scale Deep Neural Network

Abstract

In many computer vision tasks, the non-uniform blind deblurring is a serious issue. Camera shake, multiple object motions and scene depth variation are some of the major causes of motion blur which need to be tackled carefully. Video deblurring is an algorithmic approach which takes multiple corrupt, blurred and mis-focused images and estimates the clean images. In the literature, video based methodologies leverage the information available across neighbouring frames to deblur the images, which is in contrast to single image deblurring techniques. This is the reason that the performance of the video deblurring techniques depends on alignment of neighbouring frames. However, aligning neighbouring frames is a expensive task in terms of computation and it requires high level scene understanding so that one can clearly distinguish between regions that have been precisely aligned from those that have been not. Also, it is an important that any video deblurring methods should have minimum latency in real time. In this work, we gradually restore sharp video frame from a blurred one in a ‘course-to-fine' manner. We use a multi-scale deep learning based model for deblurring an image in an end-to-end manner. The experimental evaluation shows that the proposed technique performs faster when compared to other video deblurring technique with substantial improvement in qualitative and quantitative performance.