Dictionary Replacement for Single Image Restoration of 3D Scenes.

Nimisha T M, Arun Mathamkode and Rajagopalan Ambasamudram

Abstract

In this paper, we address the problem of jointly estimating the latent image and the depth/blur map from a single space-variantly blurred image using dictionary learning. The approach taken is based on the central idea of dictionary replacement viz. the sparse representation of a blurred image over a blurred dictionary is equivalent to that over a clean dictionary. While most of the dictionary-based deblurring methods consider planar scenes with space-invariant blur, we handle 3D scenes with space-variant blur caused by either camera motion or optical defocus. For a given blurred image, the dictionary blurred with the corresponding blur kernel provides the best representation with the least error. We formulate our problem of blur map and latent image estimation as a multi-label MRF and solve it using graph-cut. Experimental results on defocus as well as motion blur depict the effectiveness of our scheme.

Session

Posters 1

Files

Extended Abstract (PDF, 2M)
Paper (PDF, 8M)
Supplemental Materials (PDF, 15M)

DOI

10.5244/C.30.32
https://dx.doi.org/10.5244/C.30.32

Citation

Bibtex

        @inproceedings{BMVC2016_32,
        	title={Dictionary Replacement for Single Image Restoration of 3D Scenes.},
        	author={Nimisha T M, Arun Mathamkode and Rajagopalan Ambasamudram},
        	year={2016},
        	month={September},
        	pages={32.1-32.12},
        	articleno={32},
        	numpages={12},
        	booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
        	publisher={BMVA Press},
        	editor={Richard C. Wilson, Edwin R. Hancock and William A. P. Smith},
        	doi={10.5244/C.30.32},
        	isbn={1-901725-59-6},
        	url={https://dx.doi.org/10.5244/C.30.32}
        }