Abstract:Image colorization is the task to color a grayscale image with limited color cues. In this work, we present a novel method to perform image colorization using sparse representation. Our method first trains an over-complete dictionary in YUV color space. Then taking a grayscale image and a small subset of color pixels as inputs, our method colorizes overlapping image patches via sparse representation; it is achieved by seeking sparse representations of patches that are consistent with both the grayscale image and the color pixels. After that, we aggregate the colorized patches with weights to get an intermediate result. This process iterates until the image is properly colorized. Experimental results show that our method leads to high-quality colorizations with small number of given color pixels. To demonstrate one of the applications of the proposed method, we apply it to transfer the color of one image onto another to obtain a visually pleasing image.

Image colorization using sparse representation