Improving Image Compression Ratio Using the mUMamba Image Restoration Network

Every day, roughly three billion images travel across the digital world, moving from phones to servers and back again. This flood of visual data demands vast amounts of storage space and powerful internet connections to move without delay. To keep this system running, engineers have long relied on compression, a process that shrinks files by removing information the human eye is unlikely to notice. Traditional methods work by applying a uniform set of rules to an entire image, treating a smooth sky and a complex forest with the same mathematical approach. While effective, these standard techniques often struggle to balance the need for small file sizes with the desire for sharp, clear pictures. They can miss the subtle differences between a simple background and a detailed foreground, leading to either bloated files or blurry results.

A new approach seeks to solve this by treating different parts of an image differently, much like a librarian who knows which books are rarely read and which are checked out daily. Researchers have developed a system that identifies which parts of a photo are essential and which can be temporarily set aside to save space. The goal is to delete the unnecessary parts during transmission and then reconstruct them when the image arrives at its destination. This requires a decoder that is not just a passive receiver, but an active restorer capable of guessing the missing details with high accuracy, though not perfectly, as some reconstruction error remains.

In a recent study, researchers from Semnan University in Iran introduced a system called mUMamba to tackle this challenge. Their method combines a smart analysis of image content with a powerful new type of artificial intelligence designed for restoration. The process begins by breaking an image down into small square blocks. The system then examines these blocks at three different levels of detail, looking at them as large 16-by-16 squares, medium 8-by-8 squares, and small 4-by-4 squares. At each level, the system decides if a block is simple enough to be removed. If a block contains a smooth, unchanging area, it is marked for deletion. If it contains a sharp edge or a complex pattern, it is kept. This hierarchical approach allows the system to be far more aggressive with simple areas while carefully preserving the intricate details that define the image.

Once the removable blocks are identified, the system does not simply discard them. Instead, it sends a small amount of "clues" along with the remaining image data. For blocks that were removed because they were smooth, the system sends a simple arrow indicating the direction of the color change. For blocks removed because they contained an edge, it sends a map showing where that edge was located. Crucially, the parts of the image that were kept are not left uncompressed; they are encoded using standard JPEG compression before transmission. The original image is now much smaller, but it is incomplete. The missing pieces must be reconstructed by the receiving device using only the remaining JPEG-compressed pixels and these small clues.

To perform this reconstruction, the researchers employed a neural network, a type of computer program inspired by the human brain, built on a structure known as a U-Net. This network is shaped like a U, with a path that compresses the image information and another path that expands it back to its original size. What makes mUMamba unique is the engine it uses to process the deep layers of this network. Instead of the standard tools used in most image software, the researchers integrated a component based on a new architecture called Mamba. This component is exceptionally good at understanding long-range connections within data, allowing it to look far across the image to find the context needed to fill in the gaps. It acts like a highly skilled artist who can look at the edges of a missing puzzle piece and instantly know what the center should look like based on the surrounding picture.

The results of this approach were tested on a large collection of standard test images. The system achieved a compression ratio of 46.5 percent, indicating a significant reduction in data volume, while maintaining a high level of visual quality. When measured by standard metrics that compare the restored image to the original, the system achieved a score that indicates a very close match, with a PSNR of 25.15 dB. More importantly, the system was significantly faster and required far less computing power than other advanced methods. It used only 8.5 million parameters, which are the internal settings the computer learns during training, compared to over 40 million in competing systems. This efficiency means the system can run on smaller devices, such as mobile phones or sensors in remote locations, without needing a powerful supercomputer.

The researchers also compared their method against the current best standards used in the industry, such as BPG and VVC. They found that mUMamba could achieve the same visual quality at a lower data rate, effectively saving 18.2 percent of the bandwidth required by those standards. In direct comparisons with other modern AI-based compression tools, the new system improved the structural similarity of the images by nearly 23 percent and the clarity by 7 percent. These gains were achieved while also cutting the time it takes to process an image by nearly a third.

The study highlights that the key to this success was not just the use of a powerful neural network, but the specific way it was combined with the classical method of removing blocks. By separating the task into two distinct parts—first, the logical decision of what to remove and compress using standard JPEG, and second, the creative task of rebuilding the missing parts—the researchers created a system that is both efficient and effective. The hierarchical strategy of checking blocks at different sizes allowed the system to find more opportunities for compression without sacrificing the quality of the final picture.

While the system shows great promise, the researchers note that it currently works best with images of a fixed size and relies on standard compression for the parts of the image that are kept. Future work will aim to make the system more flexible for images of any size and to replace the remaining standard components with fully AI-driven ones. For now, the mUMamba framework stands as a significant step forward in making image transmission faster and more efficient, proving that by understanding the content of an image deeply, we can send more with less.