Image Colorization Based on Texture by Using of CNN
Jingbei Li, Huaxin Xiao, Diaoyin Tan, Maojun Zhang, Yu Liu · 2019
Colorization is a challenging task of adding colors to a given grayscale image. Related palettes can keep better consistency of color in an image and provide sharper results. This paper presents a novel technique that based on input's uniform LBP texture to generate multiple color palettes, while other methods based on reference images, words, or other subjective elements. By combing the generated color palettes, a given grayscale image can be colorized. The proposed Texture2Color model consists of two convolutional neural networks: the texture-to-palette generation networks and the palette-based colorization networks. The former network adopts the visual features and texture of the given gray image to generate colorful palettes. The latter one blends the generated palettes and grayscale images to get a plausible colorful image. Compared with the basic palette generation benchmark, the proposed method can provide a more realistic palette and achieve better performance.