Wavelet-Diffusion Collaborative Framework for Ancient Mural Restoration

Ancient wall paintings are fragile witnesses to history, capturing centuries of artistic evolution, social customs, and spiritual beliefs. From the bustling river scenes of Chinese masterpieces to the ethereal landscapes of mountain retreats, these works carry irreplaceable value. Yet, time and environment are relentless enemies. Humidity, light, pollution, and the natural aging of materials like silk and mineral pigments cause these murals to fade, crack, and crumble. Statistics indicate that a significant portion of China's estimated two million ancient paintings and calligraphy works show varying degrees of damage, creating an urgent need for restoration. For decades, the only solution was the painstaking work of human experts, a process that is slow, subjective, and risks causing further harm to the delicate surfaces. In recent years, computers have offered a new path, using artificial intelligence to fill in missing pieces and repair colors without touching the original art. However, current computer methods face a difficult trade-off: they can either reconstruct the broad shapes of a painting or restore the fine details of brushstrokes, but rarely both at once, especially when the damage is severe.

A team of researchers has developed a new approach to solve this problem, creating a system that restores ancient murals with both structural accuracy and fine textural detail. Their method, described in a recent study, combines two powerful ideas: the way computers generate new images and the mathematical technique of breaking an image into different layers of frequency. Instead of asking a computer to fix a damaged painting all at once, as most current systems do, this new framework first separates the image into its basic components. It pulls apart the smooth, broad shapes and colors from the sharp, intricate lines and textures. This separation allows the computer to work on the large-scale structure and the tiny details in parallel, rather than trying to juggle them simultaneously. By doing the heavy lifting on these separated layers, the system becomes much faster and requires far less computing power, making it possible to restore high-resolution images that were previously too difficult to process.

The core of this innovation is a collaborative framework that operates in a "wavelet domain," a technical term for a way of viewing an image based on its frequency components rather than its raw pixels. Imagine an image not as a grid of colored dots, but as a collection of different layers: one layer holding the overall shape and color, and other layers holding the edges, cracks, and fine brushwork. The researchers' system first splits the damaged mural into these four distinct layers. It then uses a generative model, a type of artificial intelligence known for its ability to create realistic images from scratch, to repair the missing parts in each layer. Crucially, the system uses the remaining, undamaged parts of the mural as a guide to ensure the new content fits perfectly with the original style. Because the computer is working on these smaller, separated layers rather than the full, massive image, it can perform the complex calculations much more quickly. Once the missing pieces in each layer are filled in, the system stitches them back together to form a complete, restored painting.

The results of this method are striking. When tested on a large collection of Dunhuang murals, which feature Buddhist scriptures and celestial scenes, the new system outperformed all existing methods. It achieved a high level of accuracy in reconstructing the image, with a score of 32.04 dB in one key measurement and a structural similarity score of 0.922, both of which are significantly better than previous techniques. In another test on a different dataset of mural images, it again achieved the best overall results, with a score of 18.95 dB. Beyond the numbers, the visual quality of the restored images is superior. Where older methods often produced blurry edges or lost the grainy texture of ancient paint, this new approach preserved the sharpness of facial contours and the subtle gradations of color. It successfully reconstructed large missing areas without distorting the composition or introducing unnatural patterns. The system proved particularly effective at maintaining the authentic artistic style, avoiding the color shifts and repetitive textures that often plague other AI restoration tools.

The researchers also tested their system against a standard version of the same technology that works directly on the full image without separating the layers. They found that the new method was not only more accurate but also dramatically faster, completing the restoration in about 9.5 seconds compared to over 35 seconds for the standard approach. This speed increase comes without sacrificing quality; in fact, the separation of layers allowed the system to focus on specific problems, such as fixing broken lines or restoring faded colors, with greater precision. The study confirms that by changing how the computer "sees" the image—breaking it down into frequency layers before repairing it—restorers can achieve a level of detail and structural integrity that was previously out of reach. This advancement offers a practical and efficient tool for preserving cultural heritage, ensuring that the intricate beauty of ancient murals can be recovered and protected for future generations.