Enhanced aero-engine blade defect detection network: MESFR-DETR with multi-scale edge selection and feature reconstruction
This paper proposes MESFR-DETR, an enhanced real-time detection transformer network featuring multi-scale edge selection and feature reconstruction modules to effectively address the challenges of detecting small, weak-feature defects in aero-engine blades, achieving superior accuracy and efficiency on both synthetic and real-world 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
The heart of a modern aircraft is its engine, a marvel of engineering where extreme heat and immense pressure are harnessed to generate thrust. Within this system, the turbine blades are the most critical components, spinning at dizzying speeds to compress air and drive the engine forward. Because these blades operate in such a hostile environment, they are constantly battered by heat, vibration, and foreign objects, leading to tiny imperfections like scratches, dents, or corrosion. Even a microscopic flaw can grow into a catastrophic failure, threatening the safety of the entire flight. For decades, ensuring these blades are safe has relied on human inspectors, who use specialized cameras and their own trained eyes to scan every surface. While effective, this manual process is slow, inconsistent, and prone to human error, especially when defects are too small or faint to be easily seen.
To solve this, researchers have turned to computer vision, teaching machines to "see" defects the way a human inspector does. However, standard computer vision tools often struggle with the unique challenges of engine blades. They tend to lose the fine, high-frequency details—like the sharp edge of a scratch or the subtle shadow of a dent—when processing images, much like how a low-resolution photo blurs out fine textures. Furthermore, many existing systems rely on complex, multi-step processes that are slow and difficult to deploy in real-world factories. A team of researchers at the Harbin Institute of Technology has developed a new approach to bridge this gap, creating a system that is both highly accurate and fast enough for real-time inspection.
The researchers, led by Debao Wei and his colleagues, introduced a new detection network called MESFR-DETR. This system is designed specifically to find the smallest, most elusive defects on aero-engine blades. Unlike older methods that might miss a tiny scratch because it blends into the background, this new network is built to pay extra attention to edges and fine details. The core of their innovation lies in three specific improvements to how the computer processes the image. First, they added a module that acts like a sharpening filter, dynamically enhancing the edges of the blade and any defects on it. This ensures that even the faintest boundary between a healthy surface and a scratch is clearly defined before the computer tries to identify it.
Second, the team redesigned the part of the system that looks for patterns across the entire image. Instead of just looking at local spots, this new module understands the relationship between different parts of the blade simultaneously. It also incorporates a technique that analyzes the image in terms of frequency, allowing the system to distinguish between the smooth, low-frequency background of the metal and the high-frequency, jagged signals of a defect. This dual approach helps the system ignore the noise of the engine's texture while zeroing in on the sharp, irregular shapes of damage. Finally, they rebuilt the network's ability to combine information from different levels of detail. By keeping a high-resolution view of the blade alongside the broader, lower-resolution view, the system can detect tiny defects that would otherwise be lost as the image is processed through deeper layers of the network.
To test their creation, the researchers needed a vast amount of data, but collecting thousands of real-world images of damaged blades is difficult and time-consuming. Instead, they generated a synthetic dataset called BladeSynth-MD, which contains 5,000 computer-generated images of blades. These images were created to look photorealistic, featuring a variety of lighting conditions, angles, and, crucially, multiple types of defects occurring on the same blade at the same time. This setup mimics the complex reality of an engine inspection, where a single blade might have a dent, a scratch, and a patch of corrosion all at once. The system was trained and tested on this dataset, and the results were striking.
The new MESFR-DETR system achieved a detection accuracy of 65.0% across all defect types, a significant improvement over the previous standard model it was based on. More importantly, it excelled at finding small defects, which are the hardest to spot. It improved the detection of these tiny flaws by 4.4% compared to the baseline. The system also proved to be highly efficient; despite its advanced capabilities, it uses fewer parameters—essentially a measure of its size and complexity—than many other leading models, making it faster and easier to run on standard hardware. When tested on a separate set of real-world images from actual in-service engines, the system maintained its high performance, successfully identifying defects that other models missed or mislabeled.
The researchers found that their approach was particularly effective against specific types of damage. For instance, it was much better at spotting dents, which are often just subtle depressions in the metal with very low contrast, and scratches, which are thin, high-frequency lines. The system's ability to preserve edge information allowed it to see these features clearly, whereas other models often treated them as background noise. The study confirms that by explicitly teaching the computer to value edge details and by reconstructing fine-grained features from different scales, it is possible to build a detector that is both sensitive enough to find the smallest flaws and fast enough to be used in a busy manufacturing or maintenance line. This work suggests a path forward where automated systems can provide a level of consistency and reliability that human inspectors alone cannot match, potentially preventing catastrophic failures before they ever happen.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.