Image Recoloring Induced by Palette Color Associations

Keywords: Image recoloring, color segmentation, image pyramid

Abstract

In this paper we present a non-interactive method for recoloring a destination image according to the color scheme found in a source image. The approach is motivated by trying to invert the working process employed in oil painting, and results are demonstrated by application to several well-known oil paintings. The algorithm uses several color models, but leans most heavily on the L-alpha-beta color space. We first color segment each image bottom-up by iteratively merging groups of pixels into connected regions of similar color. During color segmentation, a color "texture" tree is generated and associated to each region. Next, we construct classes of regions by compensating for color duplication and color similarity within the set of averaged color values obtained from regions. We extract a color palette for each image by choosing the colors of canonical region representatives from these classes. Once this palette is constructed for each image, any inverse map from the set of destination palette colors to the set of source palette colors induces a forward map from the classes of regions in the source image to sets of classes of regions in the destination image. For each source class in the range of the inverse map we transfer color from its canonical region representative to each of the associated destination regions. Color transfer occurs at the level of pixels, and uses the color texture trees associated to the regions. Our recoloring method attempts to maintain the destination image's original value structure. This is accomplished by transferring only the alpha and beta channels from the source. To make our method computationally tractable, we work within an image pyramid, transferring color layer by layer.