This paper describes a method of improving the quality of the color in color images by colorizing them. In particular, color quality may suffer from improper white balance and other factors such as inadequate camera characterization. Colorization generally refers to the problem of turning a luminance image into a realistic looking color image and impressive results have been reported in the computer vision literature. Based on the assumption that if colorization can successfully predict colors from luminance data alone then it should certainly be able to predict colors from color data, the proposed method employs colorization to 'color' color images. Tests show that the proposed method quite effectively removes color casts—including spatially varying color casts—created by changes in the illumination. The colorization method itself is based on training a deep neural network to learn the connection between the colors in an improperly balanced image and those in a properly balanced one. Unlike many traditional white-balance methods, the proposed method is image-in-image-out and does not explicitly estimate the chromaticity of the illumination nor apply a von-Kries-type adaptation step. The colorization method is also spatially varying and so handles spatially varying illumination conditions without further modification.
Ligeng Zhu, Brian Funt, "Colorizing Color Images" in Proc. IS&T Int’l. Symp. on Electronic Imaging: Human Vision and Electronic Imaging, 2018, pp 1 - 6, https://doi.org/10.2352/ISSN.2470-1173.2018.14.HVEI-541