Local affinity-based color propagation of images

Scribble-based colorization algorithms have become greatly popular due to the ease of user interaction. The key of propagating the user scribbles to the entire image is to define a similarity measure between pixels. We propose here a new affinity function with several computation times of similarity weights for each pair of adjacent pixels, while the traditional methods only with one such computation. It is proved that our method can better capture the local intrinsic structure of pixels. Furthermore, we apply a superpixel representation of the target image to reduce the computation cost. Experiments show that the new method can yields better colorization results and remarkably speed up the colorization process.