High Schooler Sparks AI for Ancient Art Restoration

Zheng Zheng's journey with art began when he started painting at just six years old. Today, as an assistant teaching professor at Northeastern University’s College of Engineering, he is channeling that lifelong passion into a new frontier: using artificial intelligence to restore historic murals.

Remarkably, the spark for this innovative project came not from a seasoned academic, but from an ambitious teenager. Qianhao Han, then a 10th-grade student in Toronto, sent a cold email to Zheng, expressing her desire to get involved in university-level research. After reaching out to several professors without success, she found a receptive audience in Zheng, whose lab at Northeastern’s Toronto campus saw the immense potential in her drive.

It was a perfect match. Han possessed a rare blend of passion for both art and science, with experience in robotics competitions and a self-taught foundation in AI. What she needed, she said, was “more professional guidance.”

After Han impressed him with a presentation on an AI project to detect pedestrians, Zheng welcomed her into his lab. Soon, she was collaborating with graduate students like Jincheng Jiang on a project she herself initiated.

“Using AI to build a tool for art restorers is actually something that I proposed,” Han says.

Introducing MurFact AI

The resulting program, named MurFact AI, is designed to analyze a mural, identify areas that are missing or distorted, and then generate a historically and artistically coherent interpretation of the complete image. The project is a testament to the shared love for history that both Han and Zheng possess, with a particular focus on the ancient murals of China.

The Unique Challenges of Restoring Murals

Restoring murals presents a unique set of challenges that go far beyond fixing a simple photograph. Zheng explains that the sheer scale, with some works stretching over 10 meters, is a significant factor. Furthermore, many murals are exposed to the elements, suffering damage from humidity, sunlight, and human interference.

The AI must also account for the three-dimensional nature of these artworks. Unlike a flat image, murals have texture and visible brush strokes that the program must consider. It needs to intelligently distinguish between different types of damage, from a jagged hole to a section bleached by the sun.

Crucially, cultural context is paramount. Zheng uses the example of a damaged mural of a Buddha with a missing hand. “We need to know some of the history about this Buddha, and then to infer what kind of gesture that Buddha may have,” he explains. The AI must therefore be trained not just on visual data, but on the cultural, social, and religious ideas of the era, some of which may be over 1,000 years old.

To achieve this level of cultural fluency, MurFact is built on an architecture similar to modern large language models like ChatGPT. It learns by having human experts label thousands of images, allowing it to understand the “grammar” of the art and make informed, semantic judgments.

A Tool for Experts Not a Replacement

Zheng is clear that the goal of MurFact is not to replace human experts. Traditional restoration is a painstaking process that can take years, involving thousands of critical decisions. He envisions MurFact as a powerful app that can assist restorationists in their work.

“He sees a world where MurFact can ‘help somehow speed up or enhance or preserve all these invaluable artifacts more efficiently.’”

This groundbreaking project, born from the ambition of a high school student, perfectly “reflects our Northeastern spirit, giving back to our community and offering opportunities,” Zheng notes.

For Han, the experience has been transformative. “It was a really good way for me to see, ‘Oh, I really enjoy this, and I want to pursue this in the future,’” she says.