AFCC: Adaptive Fusion of Local Contour Descriptors and Hierarchical Convolutional Features for Weld Formation Classification on Curved Workpieces
This paper proposes AFCC, an adaptive fusion framework that combines local contour descriptors with hierarchical convolutional features and reliability-guided weighting to achieve high-accuracy classification of weld formations on curved workpieces, significantly outperforming baseline deep learning methods in an industrial setting.
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
In the world of heavy manufacturing, the strength of a structure often depends on the invisible seams where metal pieces are joined. When workers weld metal racks, pipes, or frames, they create a raised ridge of molten metal called a bead. If this bead is perfect, the joint is strong. If it is too small, the connection is weak; if it is too large or messy, it can interfere with assembly or indicate instability. For decades, humans have inspected these joints by eye, looking for the smooth, continuous lines of a good weld versus the jagged or incomplete lines of a bad one. However, as factories produce more complex parts, human inspection cannot keep up. The challenge becomes even harder when the metal being welded is not flat but curved, like the corner of a large rack. On these curved surfaces, light reflects unevenly, and the angle of view changes constantly, making a perfect weld look different in every photograph. A computer trying to "see" these defects must learn to ignore the confusing glare and shadows to find the true shape of the weld underneath.
Researchers at National Chin-Yi University of Technology have developed a new system called AFCC to solve this specific problem. Their goal was to create an automated visual inspector that could reliably sort welds on curved metal racks into three categories: normal, excessive or distorted, and insufficient. To do this, they did not rely on a single method of looking at the image. Instead, they built a framework that mimics how a human might look at a weld from two different angles at once. One part of the system focuses on the outline and shape of the weld bead, much like tracing the edge of a shadow to understand its form. The other part looks at the texture and color details, similar to how a person might feel the roughness of a surface or notice a change in its shine. By combining these two perspectives, the system can see past the confusing reflections that often fool simpler cameras.
The researchers tested their system on a real-world dataset collected from an active metal-rack production facility. They gathered 1,586 images of welds taken directly from the factory floor, where lighting is not perfect and surfaces are often dirty with oil or metal filings. Of these images, 1,188 showed normal welds, while the rest showed defects: 218 were excessive or distorted, and 180 were insufficient. To ensure the system was truly learning and not just memorizing, they held back a specific group of 477 images to test on only after the system was fully trained. The results were striking. The new system correctly classified the welds with an accuracy of 95.39%. More importantly, it was exceptionally good at catching the defects. It missed only four of the 120 defective welds in the test set. In comparison, a standard, widely used computer vision model that was fine-tuned for this task missed 14 defective welds, and another advanced method missed eight.
The success of the system comes from how it decides which information to trust. The researchers found that the "outline" view and the "texture" view were not equally reliable in every situation. Sometimes the outline was clear but the texture was confusing due to glare; other times, the texture was clear but the outline was broken by noise. Instead of giving both views equal importance, the system learned to weigh them based on how well each performed during its training. It assigned more influence to the view that was proving more accurate for a specific type of defect. This adaptive approach allowed the system to ignore the weaker signals and focus on the strongest evidence, leading to a much higher success rate than if it had simply averaged the two views together.
Beyond just sorting the parts, the system offers a practical path forward for factory automation. Because it can distinguish between three specific conditions, it can direct each weld to the correct next step. A normal weld can be sent straight through. A weld with too much material can be routed to a grinding station to be smoothed down. A weld with too little material can be sent for a second pass of welding to fill the gap. The researchers also explored a way to make this process even more efficient by using artificial intelligence to explain why a weld was flagged as defective. They found that by only asking the AI to write an explanation when a defect was actually found, rather than for every single image, they could reduce the number of computer requests by nearly 73%. This saved significant time and computing power while still covering almost all the defective parts.
The study highlights that while modern computers are powerful, they still struggle with the messy reality of curved surfaces and uneven lighting. The researchers demonstrated that combining simple geometric measurements with complex visual patterns creates a more robust solution than relying on deep learning alone. They also noted that the system is not perfect; it can sometimes be fooled by small bits of metal debris, known as spatter, that sit near the weld and look like an extra ridge. However, by identifying these specific failure modes, the researchers have a clear path for future improvements. The work proves that for curved industrial parts, a hybrid approach that respects both the shape and the texture of the weld is the most reliable way to ensure quality, offering a clear, data-driven method to keep production lines moving safely and efficiently.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.