2016/12/07 by Seungjun Nah, Nah, Seungjun, Tae Hyun Kim +3 · 75 citations
Computer Science · Engineering · #Advanced Image Processing Techniques #Image Processing Techniques and Applications #Advanced Vision and Imaging
paper · pdf · doi:10.48550/arxiv.1612.02177
Non-uniform blind deblurring for general dynamic scenes is a challenging\ncomputer vision problem as blurs arise not only from multiple object motions\nbut also from camera shake, scene depth variation. To remove these complicated\nmotion blurs, conventional energy optimization based methods rely on simple\nassumptions such that blur kernel is partially uniform or locally linear.\nMoreover, recent machine learning based methods also depend on synthetic blur\ndatasets generated under these assumptions. This makes conventional deblurring\nmethods fail to remove blurs where blur kernel is difficult to approximate or\nparameterize (e.g. object motion boundaries). In this work, we propose a\nmulti-scale convolutional neural network that restores sharp images in an\nend-to-end manner where blur is caused by various sources. Together, we present\nmulti-scale loss function that mimics conventional coarse-to-fine approaches.\nFurthermore, we propose a new large-scale dataset that provides pairs of\nrealistic blurry image and the corresponding ground truth sharp image that are\nobtained by a high-speed camera. With the proposed model trained on this\ndataset, we demonstrate empirically that our method achieves the\nstate-of-the-art performance in dynamic scene deblurring not only\nqualitatively, but also quantitatively.\n