Transformer Image Dehazing Method Based on Physical Prior Guidance and Adaptive Feature Modulation
This paper proposes a Transformer-based single-image dehazing network that integrates physical prior guidance with adaptive feature modulation to effectively handle spatially non-uniform haze while preserving fine structural details, achieving state-of-the-art performance on multiple benchmark datasets.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
When a thick fog rolls in, the world does not simply turn gray; it turns flat. The air fills with tiny water droplets or dust that scatter light, washing out the deep blues of the sky and turning sharp edges into soft, indistinct smudges. For a human, this is a temporary inconvenience, but for a machine trying to see—like a self-driving car navigating a highway or a surveillance camera monitoring a city street—it is a fundamental failure. The machine loses the ability to tell where a road ends and a building begins, or to distinguish a pedestrian from a shadow. Scientists have long tried to teach computers to reverse this process, to strip away the haze and reveal the clear scene hidden beneath. Early attempts relied on strict rules about how light behaves in the atmosphere, while newer methods use artificial intelligence to learn from thousands of examples. However, these learning systems often struggle when the fog is uneven or the lighting is tricky, sometimes smoothing out important details or inventing colors that were never there.
A team of researchers at Southeast University has developed a new approach that bridges the gap between these two worlds. They created a system that combines the flexibility of modern artificial intelligence with the reliable logic of physics. Instead of letting the computer guess how to clear the image, they guide it with a map of where the fog is thickest. This map, known as a transmission map, is an estimate of how much light is blocked at every point in the picture. By feeding this physical clue directly into the learning process, the system knows exactly where to focus its efforts. It does not just try to make the picture brighter; it learns to restore the specific structures that the fog has hidden, ensuring that a distant tree remains a tree and a nearby fence stays sharp.
The core of their invention is a network that mimics the way a human might look at a foggy scene: first taking in the broad view, then zooming in on the details. The researchers built a multi-layered structure that processes the image at different scales simultaneously. At the broadest level, the system uses a type of advanced attention mechanism to understand the relationship between distant parts of the image, recognizing that fog often covers a whole landscape in a connected way. At the same time, it uses specialized filters to keep an eye on the tiny, local details that define edges and textures. This dual approach allows the system to maintain the big picture while refusing to lose the fine lines that make an image recognizable.
To make this work even better, the team introduced a dynamic way for the system to adjust its own behavior. Imagine a camera lens that can instantly change its focus depending on whether it is looking at a smooth wall or a complex pattern of leaves. In this system, the computer analyzes the specific type of haze it is facing and changes how it processes the information. If the fog is light and the details are clear, it focuses on preserving those fine textures. If the fog is heavy and the scene is washed out, it shifts its attention to the broader context to make sense of the shapes. This adaptability is guided by the physical map mentioned earlier, which acts as a constant reminder of where the atmosphere is most dense. The system also pays special attention to edges, using a technique that highlights the boundaries between objects, ensuring that the fog does not blur the lines that separate a car from the road or a person from a building.
The researchers tested their method on a wide variety of scenes, including both indoor environments like offices and outdoor settings like city streets and rural roads. They compared their results against several other leading methods, including older rule-based techniques and newer artificial intelligence models. The results showed that their approach consistently produced clearer images with higher structural accuracy. In tests on outdoor scenes, their method achieved a score of 24.57 on a standard measure of image quality, while preserving the structure of the image with a score of 0.907. In indoor tests, the performance was even stronger, reaching 28.78 for quality and 0.928 for structure. These numbers indicate that the system not only removes the gray haze but also keeps the sharpness and arrangement of objects intact, a feat that many other methods struggle to achieve simultaneously.
Perhaps most importantly, the system proved robust when the fog became extremely thick. As the haze density increased, other methods tended to lose their ability to distinguish shapes, often resulting in blurry or distorted images. The new method, however, degraded much more gracefully. Even under heavy fog conditions, it maintained a high level of structural integrity, keeping the outlines of objects clear and the colors true. This suggests that the combination of physical guidance and adaptive learning creates a system that is less likely to make mistakes when the visual information is scarce. The researchers found that each part of their system played a unique role: the physical map told the system where to look, the attention mechanisms helped it understand the scene's layout, and the dynamic adjustments allowed it to fine-tune its response to the specific conditions.
While the results are promising, the researchers are careful to note that their tests were conducted using computer-generated fog paired with clear images. This setup allows for precise measurement of how well the system works, but real-world fog can be more complex, containing variations in light and particle density that are hard to simulate. The team acknowledges that further testing on real-world footage is needed to confirm that the method works perfectly in uncontrolled environments. They also point out that the system requires more computing power than some simpler models, which could be a challenge for devices with limited resources. Despite these limitations, the work offers a significant step forward in how machines perceive a degraded world. By teaching artificial intelligence to respect the laws of physics while learning from data, the researchers have created a tool that sees through the fog not just by guessing, but by understanding.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.