Light-sensitive camera sensors that can record images even in very dark conditions usually also use the near infrared spectrum (NIR). Similarly, special night vision systems can detect infrared light and convert it into images that can be perceived by humans. However, images taken in this way suffer from the fact that - similar to the human eye at night - they no longer show the original colors, but only pixels differentiated by their brightness. For this reason, scientists at the University of California have now developed a deep-learning algorithm that can be used to transform monochromatic images taken by an infrared camera into colored photos even in complete darkness (plus infrared light).
Face in visible light and reconstruction of colors by neural network from infrared image only
The new imaging algorithm, based on optimized deep-learning architectures, can use the infrared image of a scene to predict the scene&s appearance in the visible spectrum and render it as if it were illuminated by light in the spectrum visible to humans. Thus, a scene could be digitally rendered in the visible spectrum for humans, i.e., in color, even if it is otherwise in complete darkness and illuminated only by infrared light.
To this end, the researchers used a monochromatic camera sensitive to visible and near-infrared light to capture an image dataset of printed images of faces under multispectral illumination that includes standard visible wavelengths of red (604 nm), green (529 nm) and blue (447 nm), as well as infrared (at 718, 777 and 807 nm). A neural network was then trained to predict visible spectrum images from near-infrared images only.
Object illuminated by visible and infrared light.
The new algorithm was indeed able to predict visible spectrum images from near infrared information only. However, the neural network trained in the research work only with human faces. In order to be able to correctly colorize scenes with other objects, such scenes would have to be trained accordingly in the future. Then this method would be a perfect complement, for example, for the new AI method (recently presented) by us www.slashcam.de/news/single/Nur-Sternenlicht-als-Beleuchtung--Neuer-KI-Algorit-17213.html, which is able to remove the image noise almost completely from videos recorded under extremely weak illumination of fractions of a lux.