Hierarchical Reinforcement Learning-Based Adaptive Denoising and Reversible Color Restoration Method for Complex Degraded Murals

Murals are more than just paintings on a wall; they are fragile records of human history, often clinging to ancient cave surfaces or crumbling plaster. Unlike a canvas in a museum, these artworks have spent centuries exposed to the elements, suffering from a unique and chaotic mix of damage. They are not merely faded or dirty; they are often covered in a complex layering of noise, cracks, stains, and peeling pigment, where the very texture of the paint has begun to weaken. Restoring them is a delicate balancing act. If a restorer cleans too aggressively, they might erase the fine lines of a brushstroke or the subtle history of aging. If they are too cautious, the image remains obscured. The goal is to reveal the original art without inventing new details or permanently altering what remains, a process that requires a deep understanding of where to touch and how hard to press.

For years, computer scientists have tried to automate this delicate work using artificial intelligence. Early attempts often treated the entire image the same way, applying a single set of rules to clean up the whole picture at once. This approach works well for simple photos, but it fails with murals because the damage is rarely uniform. A face might be cracked while the background is merely dusty, and a bright spot of noise might look exactly like a decorative pattern. Applying a heavy-handed cleaning filter to the whole image can smooth out the cracks but also blur the face, destroying the very details that matter most. The challenge, therefore, has been to create a system that can look at a mural, decide which specific part needs help, and choose the gentlest, most appropriate tool for that specific spot, all while ensuring that the changes can be traced and reversed if necessary.

In a recent study, researchers proposed a new way to tackle this problem by teaching a computer to make decisions in steps, much like a human restorer would. Instead of forcing the computer to fix the whole image in one go, they broke the task down into two levels of thinking. The first level acts as a supervisor, scanning the entire mural to decide which area needs attention first. It looks at the damage and asks, "Is this a face that needs careful handling, or a background that can be cleaned more aggressively?" Once a region is selected, a second, more specialized level takes over. This lower level decides exactly how to fix that specific spot, choosing from a menu of actions like removing noise, adjusting color, or preserving texture. Crucially, the system is designed to be reversible, meaning it keeps a record of every change so that the original state can be recovered if the restoration goes too far.

The researchers tested this two-level system on thousands of images, including simulated murals with heavy damage and real-world examples from ancient sites. They compared their method against other powerful computer programs that try to restore images. The results showed that their approach, which they call a hierarchical reinforcement learning framework, produced clearer and more accurate results than the others. In tests, the new method improved the clarity of the images significantly, reducing the visual noise while keeping the sharp edges of the artwork intact. More importantly, it made fewer mistakes with color. While other methods often shifted the colors too far, making them look too bright or unnatural, this system kept the colors much closer to what they should be, reducing the average color error by nearly thirty percent.

The study also revealed that the way the computer makes decisions is just as important as the final picture it produces. When the researchers removed the "supervisor" level and let the computer try to fix the whole image at once, the quality dropped noticeably. The system became less able to distinguish between damage and important artistic details. Similarly, when they removed the ability to reverse changes, the computer became more aggressive, often over-correcting colors and losing the subtle texture of the paint. These findings suggest that the step-by-step decision-making process is essential for handling the messy, uneven reality of damaged art. The computer learns to be patient, focusing on one area at a time and stopping when it senses that a region is stable, rather than forcing a uniform fix across the entire surface.

Beyond just getting the numbers right, the system offers a way to control the style of the restoration based on the intended use. The researchers showed that the same tool could be tuned for different purposes. If the goal is strict historical documentation, the system can be set to make very few changes, preserving the current state of the mural even if it looks a bit faded. If the goal is a public exhibition, the system can be adjusted to make the colors more vibrant and the details sharper, helping viewers see the art more clearly. This flexibility means that digital restoration does not have to be a one-size-fits-all process. It can be a controlled, transparent operation where conservators decide how much intervention is appropriate, knowing that the digital version can always be dialed back to a safer state.

The study was conducted using a large collection of images, including thousands of paintings used to teach the computer what art looks like, and hundreds of specific mural images that were artificially damaged to test the system's limits. While the results are promising, the researchers note that the current tests were simulations. The numbers and visual comparisons were generated by running the program on these controlled datasets, which allows for precise measurement but does not yet prove how the system will perform on a real, centuries-old wall in a cave. The author suggests that the next step is to test the method on actual heritage sites and to involve human experts to judge the results. Until then, the work stands as a strong demonstration that giving computers a structured way to think about restoration—deciding where to act and how to act—can lead to better, safer, and more respectful preservation of our shared history.