Fast Image Colorization Based on Local and Global Consistency

Image colorization refers to the process of semiautomatic adding colors to a grayscale image. A novel image colorization approach based on local and global consistency is proposed, which applies the graph-based semi-supervised learning theory to solve the problem of colorization. Firstly, constructing a graph model according to the similarities between pixels, and then determine a quadratic cost function with sufficient smoothness which can fully guarantee the image structure information, finally the colors are propagated to the un-colored pixels from colored pixels by minimizing the cost function. In order to meet the time requirements when dealing with high resolution images, the authors further propose a segmentation-based approach, instead of directly on pixels, which regards image segmented region as the basic unit during colors propagating. Experiments on different natural images demonstrate that our approach´s computational complexity is significantly reduced while achieves high quality colorization results compared with two existing techniques.