Picture Color is used to modernize classic black and white films or provide artistic control over grayscale images with diverse color distribution. However, current methods often result in unsaturated and unnatural colors. A recent article on arXiv.org proposes BigColor, a novel image coloring method that synthesizes vivid and natural colors for complex structured wild images.
Image credit: arXiv: 2207.09685 [cs.CV]
The researchers exploit the spatial structure of the input grayscale image using a built-in encoder that allows magnification representative space of the generator compared to the conventional coloring process.
BigColor can handle images of arbitrary size, which is not feasible with conventional methods. Furthermore, multimodal color correction results can be synthesized using different condition vectors for the network. Extensive tests confirm that BigColor performs better than previous methods.
In order to create realistic and vivid colors, bio-based elements have recently been exploited. However, such reproduction primitives often fail for complex images in nature due to their limited representation space. In this paper, we propose BigColor, a new color approach that provides the ability to create vivid colors for diverse wild images with complex structure. Whereas primitives were previously trained to synthesize both color and image structure, we learned a composite color before focusing on color synthesis based on the spatial structure of the image. In this way, we relieve the burden of compositing the image structures from the previous one and expand its representation space to cover diverse images. To this end, we propose a BigGAN-inspired encoder generator network that uses spatial feature maps instead of spatially flattened BigGAN latent codes, resulting in a space of representations extend. Our method enables powerful coloration of diverse inputs in a single transition, supports arbitrary input resolutions, and provides multimodal color correction results. We demonstrate that BigColor is significantly superior to existing methods, especially on highly structured wild-type images.
Research articles: Kim, G., “BigColor: Colorization using preformed colors for natural images”, 2022. Link: https://arxiv.org/abs/2207.09685
