Non-uniform blind deblurring for general dynamic scenes is a challengingcomputer vision problem since blurs are caused by camera shake, scene depth aswell as multiple object motions. To remove these complicated motion blurs,conventional energy optimization based methods rely on simple assumptions suchthat blur kernel is partially uniform or locally linear. Moreover, recentmachine learning based methods also depend on synthetic blur datasets generatedunder these assumptions. This makes conventional deblurring methods fail toremove blurs where blur kernel is difficult to approximate or parameterize(e.g. object motion boundaries). In this work, we propose a multi-scaleconvolutional neural network that restores blurred images caused by varioussources in an end-to-end manner. Furthermore, we present multi-scale lossfunction that mimics conventional coarse-to-fine approaches. Moreover, wepropose a new large scale dataset that provides pairs of realistic blurry imageand the corresponding ground truth sharp image that are obtained by ahigh-speed camera. With the proposed model trained on this dataset, wedemonstrate empirically that our method achieves the state-of-the-artperformance in dynamic scene deblurring not only qualitatively, but alsoquantitatively.
展开▼