Vectorized Colorization of Icon Line Art Based on Closed Contour Extraction

In the field of icon line art colorization, several Generative Adversarial Networks (GANs) based methods have achieved remarkable success. However, these methods often suffer from issues such as noise, color inconsistencies, and distortion when the generated color icons are enlarged. To address these challenges, we propose a novel approach for vectorized colorization of icon line art (LAVC), leveraging the principle of closed contour extraction. Specifically, our contour semantic descriptor (CSD) aims to fill the vector paths of the same descriptors with the same color for color inconsistencies. Our fusion model fuses the SVG line art, contour semantic descriptor, and color raster image generated from line art, to generate high-quality color vector icons without noise and distortion. Furthermore, we collect two datasets, IconLine and ClipLine, which provide high-quality line art and color image pairs for icons. Experimental evaluations conducted on our datasets demonstrate that our method outperforms existing techniques in terms of icon line art colorization, while maintaining distortion-free scalability.