Thermal Image Colorization Using Deep Learning

The problem of coloring optical images (gray levels) has been studied extensively [1], and there are systems that can perform high-quality coloring of such images. Unlike an optical image, in which the information received by the sensor is returned light from an object, in a thermal image the information received is heat emitted from the object itself, in addition to the light reflected from it. Thermal images consist of waves in the infra-red spectrum, which is not visible. These differences between the image types make the problem of coloring thermal images a different problem.
The goal of this project is to build a dedicated deep learning system for coloring thermal images, so that the colors of the resulting image are as authentic as possible, and in addition the sharpness of the image will be maintained. The system is based on a GAN (Generative Adversarial Network) architecture, where the database consists of pairs of thermal and optical images that were taken from a public FLIR image database. The pairs of images in the database went through a pre-processing that included image registration between the pairs. The results show that the system is able to color thermal images qualitatively, and the sharpness of the image is relatively maintained.

Thermal Image Colorization Using Deep Learning