Improved IR-Colorization using Adversarial Training and Estuary Networks

Matthias Limmer and Hendrik Lensch

Abstract

This paper investigates deep learning based image translation techniques to perform a colorization of near-infrared (NIR) images. Current deep neural network design patterns like inception modules, skip-connections or multi-scalar processing are combined and examined. Adversarial training is used to improve the realism of the resulting image in the translated domain. The presented approach is trained and evaluated on NIR-RGB image pairs from a real-world dataset containing a large amount of road scene images in summer.

Session

Posters

Files

PDF iconPaper (PDF)

DOI

10.5244/C.31.82
https://dx.doi.org/10.5244/C.31.82

Citation

Matthias Limmer and Hendrik Lensch. Improved IR-Colorization using Adversarial Training and Estuary Networks. In T.K. Kim, S. Zafeiriou, G. Brostow and K. Mikolajczyk, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 82.1-82.14. BMVA Press, September 2017.

Bibtex

            @inproceedings{BMVC2017_82,
                title={Improved IR-Colorization using Adversarial Training and Estuary Networks},
                author={Matthias Limmer and Hendrik Lensch},
                year={2017},
                month={September},
                pages={82.1-82.14},
                articleno={82},
                numpages={14},
                booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
                publisher={BMVA Press},
                editor={Tae-Kyun Kim, Stefanos Zafeiriou, Gabriel Brostow and Krystian Mikolajczyk},
                doi={10.5244/C.31.82},
                isbn={1-901725-60-X},
                url={https://dx.doi.org/10.5244/C.31.82}
            }