Structure-Preserving Colorization Based on Quaternionic Phase Reconstruction

A novel semiautomatic colorization method is proposed based on quaternionic phase reconstruction. In this method, each color pixel is represented as a quaternion, whose polar magnitude and polar phase are recovered from the intensity of original grayscale image and color scribbles of user’s manual input, respectively. To conduct structure-preserving colorization, color diffusion is restrained across global image structures, which are extracted using hierarchical edge representation along with structural importance measurement. To identify local spatial relationship between neighboring pixels, Gabor wavelets are applied to compute the similarity of local phase patterns. Our method is highlighted in well preserving image structures during colorization, where the color image is acquired by solving a linearly constrained quadratic optimization problem. Specifically, we develop a method to guide the user to scribble on the monochrome image, so that effective color propagation from less manual input can be expected. Experimental results demonstrate that our colorization method outperforms the state-of-the-art method in structural preservation and relatively better colorization results are available if the proposed rule of scribble user guidance is adopted.