Near-Infrared Image Colorization with Weighted UNet++ and Auxiliary Color Enhancement GAN

Abstract

We propose a novel GAN-based method for near-infrared image colorization. This method innovatively rebalances the color of the colorization image by importing a luminance channel and a feature weight-driven color generator. We set the weighted UNet++ structure in the generator for colorization results with the detail of focal objects. A color enhancement network composed of a deeper luminance network and a colorimetric network is used for global color balance to improve the color quality of the generated color images. Our network is trained and evaluated on two datasets. According to the FID, SSIM and PSNR results, our network performs well, with good recovery effects for both overall color and detailed color and outperforming the current state-of-the-art methods.