It’s a common trope in movies and TV shows that an investigator finds a fuzzy image of the perpetrator and tells the forensic specialist or computer to “enhance,” and through a series of clicks, the image becomes crystal clear.
Those in the know would tell you that such low-resolution images can only be enhanced so much and rarely to such a point, but it makes for good television. That fiction will soon be a reality.
Draper researchers have figured out how to use generative artificial intelligence (AI) systems to create synthetic high-resolution images from low-resolution pictures—an advancement that will be instrumental in enhancing space flight and moon landings.
“We want the best-performing navigation systems for our astronauts,” said Nigel Lee, Draper’s project lead. “If we can feed more information into our navigation algorithms, they will perform better and reduce mission risk.”
Generative AI uses existing data and information to “learn” how to answer queries based on the input data. Examples of generative AI that have become popular in the wider culture include ChatGPT, which generates conversational answers to text-based prompts, and “AI art” models such as DALL-E, which generate images from text-based prompts.
Draper is doing something similar with image-to-image translation, using generative adversarial networks (GANs) to generate high-resolution output images from low-resolution input images.
By feeding the AI system crisp and blurry versions of the same images — as well as other relevant, contextual information (e.g., solar angle, altitude, etc.) — Draper researchers are teaching the AI system how to identify key markers in low-resolution images and to show it what the images should look like in a higher-resolution. The system can then apply that knowledge to enhance other blurry images.
A Cost-Effective Option to Fill in our Picture of the Moon
Despite several crewed landings and subsequent advancements in long-distance imaging, we still do not have a complete picture of the moon’s surface. The astronauts who landed there 50 years ago lacked the technology to take high-quality images by today’s standards. And current technologies capable of capturing high-resolution images from thousands of miles away are expensive.
Using generative AI to transform blurry images into crisp versions is a cost-effective solution for developing a more complete picture of the lunar surface.
Using Synthetic Images to Refine Spaceflight and Lunar Landing Systems
As a leader in guidance, navigation, and control (GNC) systems for space travel, Draper was instrumental in getting the first astronaut to the moon. Today, our engineers are enhancing GNC technologies to support the Artemis program, which will return humans to the lunar surface for the first time in 50 years.
By increasing the quantity and quality of data within the algorithms that support GNC systems, we increase the reliability and precision of the algorithms and reduce the risk of problems.
“To land on the moon or walk on the moon, there is a certain amount of risk,” said Lee. “We want to reduce that risk, and to do that, we need the best data possible.”
Imagining an Image
This internally funded project, which also involved researchers Evelyn Stump, Andrew Olguin, Andrew Sandberg, and Jena Nawfel, is just one example of the many ways Draper demonstrates our commitment to solving scientific challenges facing the nation. The project was part of Draper’s AI Incubator program, which interviewed program managers and potential clients to identify important problems that needed to be solved where AI could provide novel solutions.
Over the course of the three-month project, the team was able to demonstrate the generation of realistic-looking, high-resolution synthetic lunar images from low-resolution input images for a variety of lunar features, image resolutions, and training conditions, indicating the robustness of the method and its potential for wider use.
“This is one of the success stories of the AI incubator program,” said Lee. “For the lunar imaging problem, the team was quickly able to demonstrate the effectiveness of generative AI and its applicability to many problems of interest."