Image Color Transfer with Naturalness Constraints

摘要

In consumer imaging applications involving photo collages orrncomposition of user photos with professional artwork, inconsistentrncolor appearance of photos and artwork from different sources canrnresult in compositions that do no look aesthetically pleasing.rnUsers often express a desire to modify individual images to achieverna more consistent color appearance. Prior work in color transferrnthat extract the color properties of one image and apply it tornanother have shown very interesting results [1,2]. These worksrnfocused on achieving an artistic effect, usually without thernconstraint of conserving object color. In consumer imaging, wernhave to be more conscious about conserving general object colorrnand especially skin tones, which are not amenable to aggressiverncolor change. In this paper we describe an algorithm to estimaternthe color and tone properties of an image and transfer thesernproperties to another image under a strong naturalness constraint.rnIn our method, color changes are constrained to correspond tornincomplete adaptation under natural illuminants. We use a simplernBayesian method to characterize scene color properties, expressedrnas scene color temperature and illumination levels. An existingrncolor adaptation model RLAB [3] is used to apply color changesrnby simulating incomplete adaptation to a colored targetrnilluminant. We emphasize that this is not a method of white pointrnestimation nor a white balance procedure Rather, we use colorrnadaptation models as a means to ensure color adjustments to bern“plausible”, and therefore maintain a natural appearance to thernimages even after significant color adjustments.