Infrared Image Colorization Network using Variational AutoEncoder

This paper presents a novel method for colorizing the near-infrared (NIR) images. Recent deep learning based approaches for NIR colorization utilized an auto-encoder structure to map a single channel to the color domain. However, these methods only provide latent codes of a NIR image using an encoder structure, which simply vectorizes an input image to a latent vector. In this vectorization, all semantic information is mutually considered although each latent code represents independent information. To tackle this issue, we propose a NIR colorization network using variational auto-encoder. This network encodes a NIR image into latent codes with probabilistic distributions. Moreover, we embed a novel correlation module to interactively consider luminance and chrominance features. It facilitates the proposed network to generate better textures and color information. Our model achieves comparable performances on the large-scale dataset: video surveillance in a day (VSIAD).