Ethnic Costume Grayscale Image Coloring Method with Improved Pix2Pix

Abstract

Grayscale image coloring is challenging for high-resolution images of ethnic costume class, and current coloring methods are mainly applicable to low-resolution images, which pay less enough attention to local regions. While ethnic costume images are characterized by rich semantic information, diverse colors and high resolution, the original methods are difficult to show good results on ethnic costume dataset and prone to problems such as inaccurate coloring in local areas. In this paper, we propose a method of colorization that combining global features and attention mechanism with Pix2Pix, by injecting grayscale map into different layers of the generator as global features, and accelerating the speed of the convergence by the attention mechanism. The experimental results show that the proposed method has better coloring effect on ethnic costume images and could generate higher quality images.