Colorization using optimization

Abstract

Figure 1: Given a grayscale image marked with some color scribbles by the user (left), our algorithm produces a colorized image (middle). For reference, the original color image is shown on the right.

Discriminative Textural Features for Image and Video Colorization

“…Levin et al formulated an optimization problem [1] based on an assumption that neighboring pixels of similar intensity should have similar color values under the limitation that the colors indicated in the scribbles remain the same. Yatziv and Sapiro proposed a method [2] for determining propagation paths in the image by minimizing geodesic distances from every scribble.…”

“…Here, the image is transformed into a 49-dimensional DCT feature space, which is further reduced to 10 dimensions using linear discriminant analysis (LDA). The textural features are only used to determine location of transferred microscribbles which are subsequently propagated using Levin's method [1].…”

“…An adaptive approach presented in the Irony's color transfer [15] deals with this problem by using LDA, but as this method is based on local frequency analysis, it is required that the regions differ in the frequency domain. The obtained feature space is used only to match the most similar pixels, while the propagation is performed in the luminance domain using Levin's method [1].…”

“…We compared the proposed method with two wellestablished colorization techniques proposed by Levin [1] and Yatziv [2]. The first one is published in the form of MATLAB code and for the latter a Java applet is available to colorize a fixed set of images (for others we used our implementation).…”

“…Such information cannot be delivered by recent image understanding systems yet and therefore this process is currently intended to be human-assisted. Existing solutions make it possible to limit the necessary human actions to defining the initial color scribbles [1]- [6] or giving examples of color images of similar content. For scribble-based routines the colors are propagated from the marked image regions using distance transform in the luminance domain.…”