Colorization of Near-Infrared Images using Generative Adversarial Networks

Abstract

Near-infrared cameras can get more detail than RGB cameras even in dark environments. However, near-infrared images are black-and-white images, and RGB information is recommended when presenting information to humans. In recent years, there have been a few studies on the colorization of near-infrared images. Most of the results of previous studies are only when NIR and RGB images are paired. Therefore, in this paper, we proposed the conversion method that does not use pair images based on the cycle generative adversarial network. We performed validation experiments using two networks and evaluated by mean angle error. In addition, object detection was performed on the colorized image, and the number of detections before and after colorization was compared. As a result, it was found that it is necessary to perform the conversion without breaking the shape while maintaining the edges of the original image.