Multi-Band NIR Colorization Using Structure-Aware Network

Near infrared (NIR) image can capture the scene in the low light condition without noise unlike RGB. Therefore, it has been widely used in low light vision problems and is often fused with its RGB counterpart for RGB image enhancement. However, there can be some situations that an RGB image can be hardly captured like extremely low-light condition. To cope with this problem, researches to convert NIR image to RGB have been recently conducted based on deep learning networks. These methods show a good performance relatively, but they have some limitations of performance. In this paper, we propose a deep network to colorize multi-band NIR images to RGB using our new dataset. The proposed method attempts to exploit the correlation between individual NIR band and RGB by using multi-band NIR images. It can successfully colorize the multi-band NIR images using two-branch structure and the constraint of the proportional gradient between NIR and RGB.