Color Transfer between Images

ne of the most common tasks in image processing is to alter an image's color. Often this means removing a dominant and undesirable color cast, such as the yellow in photos taken under incandescent illumination. This article describes a method for a more general form of color correction that borrows one image's color characteristics from another. Figure 1 shows an example of this process, where we applied the colors of a sunset photograph to a daytime computer graphics rendering. We can imagine many methods for applying the colors of one image to another. Our goal is to do so with a simple algorithm, and our core strategy is to choose a suitable color space and then to apply simple operations there. When a typical three channel image is represented in any of the most well-known color spaces, there will be correlations between the different channels' values. For example, in RGB space, most pixels will have large values for the red and green channel if the blue channel is large. This implies that if we want to change the appearance of a pixel's color in a coherent way, we must modify all color channels in tandem. This complicates any color modification process. What we want is an orthogonal color space without correlations between the axes. Recently, Ruderman et al. developed a color space, called lαβ, which minimizes correlation between channels for many natural scenes. 2 This space is based on data-driven human perception research that assumes the human visual system is ideally suited for processing natural scenes. The authors discovered lαβ color space in the context of understanding the human visual system , and to our knowledge, lαβ space has never been applied otherwise or compared to other color spaces. There's little correlation between the axes in lαβ space, which lets us apply different operations in different color channels with some confidence that undesirable cross-channel artifacts won't occur. Additionally, this color space is logarithmic, which means to a first approximation that uniform changes in channel intensity tend to be equally detectable. Because our goal is to manipulate RGB images, which are often of unknown phosphor chromaticity, we first show a reasonable method of converting RGB signals to Ruderman et al.'s perception-based color space lαβ. Because lαβ is a transform of LMS cone space, we first convert the image to LMS space in two steps. The first is a conversion from RGB …