Automatic Defect Detection and Intelligent Grading System for University Ceramics Integrating Computer Vision

In the world of ceramics, the journey from raw clay to a finished, gleaming plate or tile is a delicate dance of heat and chemistry. The final step, firing, transforms the material, but it also introduces a persistent enemy: surface flaws. Tiny cracks, trapped air bubbles, or subtle shifts in color can ruin a piece, turning a potential masterpiece into scrap. For centuries, the only way to catch these defects has been the human eye. A skilled inspector scans the surface, looking for imperfections that might be as thin as a hair or as small as a pinprick. However, this method is slow, tiring, and prone to error. Fatigue sets in, eyes miss the faintest details, and the judgment of what constitutes a "defect" can vary from person to person. In university laboratories, where students and researchers craft small batches of unique, artistic, or experimental ceramics, these challenges are magnified. The pieces are often one-of-a-kind, the production runs are tiny, and the need for precise, repeatable data to guide teaching and research is high.

This is where the intersection of ancient craft and modern technology offers a new path. Researchers at Jingdezhen Ceramic University have developed a system that acts as a tireless, hyper-attentive observer for these ceramic pieces. By combining specialized cameras with advanced artificial intelligence, they have created a way to see what the human eye often misses and to grade the quality of each piece with mathematical consistency. The goal was not to replace the artist, but to provide a reliable tool that can handle the repetitive, high-stakes task of inspection, freeing up human experts to focus on the creative and educational aspects of their work. The result is a system that sees the invisible, counts the uncountable, and makes decisions with a speed and accuracy that transforms how quality control is done in a small-batch, academic setting.

The journey begins with the camera itself, which faces a unique challenge: the very thing that makes ceramics beautiful—their shiny, glass-like glaze—is the same thing that makes them hard to inspect. When light hits a glossy surface, it bounces back in a blinding glare, often hiding the very cracks and bubbles the system is trying to find. To solve this, the researchers designed a special lighting setup. They used a ring of lights that sends out polarized light, a specific type of light wave that vibrates in a single direction. When this light hits the smooth, shiny glaze, it bounces back with its direction unchanged. However, when it hits a defect like a crack or a bubble, the light scatters and loses its direction. By placing a filter in front of the camera that blocks the unchanged, blinding glare but lets the scattered light from the defects pass through, the system creates a clear, high-contrast image where flaws stand out sharply against the background. It is like wearing sunglasses that block the glare of the sun but allow you to see the details of the landscape beneath.

Once the camera captures this clear image, the computer must clean it up. The images can be slightly blurry or distorted due to the vibration of the conveyor belt or the movement of air in the kiln area. The system uses a sophisticated digital filter to smooth out the random noise without blurring the sharp edges of the defects. It then analyzes the image in a way that separates the regular, repeating patterns of the glaze texture from the irregular, chaotic patterns of a defect. This allows the computer to ignore the natural grain of the ceramic and focus entirely on the flaws. The result is a pristine, enhanced image ready for the next stage: the artificial intelligence.

The heart of the system is a deep learning model, a type of computer program trained to recognize patterns. The researchers chose a powerful architecture known as YOLO, which stands for "You Only Look Once," because it can detect objects in an image almost instantly. However, standard versions of this model struggled with the specific challenges of ceramics, such as the tiny size of some bubbles or the complex shapes of cracks. To fix this, the team added a special attention mechanism that helps the computer focus on the most important parts of the image, much like a human would focus their gaze on a specific spot. They also optimized the way the model looks at the image at different scales, ensuring it can spot a tiny pinhole just as easily as a long, jagged crack.

A significant hurdle in training such a system is the lack of data. In a university lab, they might only produce a few dozen pieces at a time, and defects are rare. This means there are very few examples of cracks or bubbles for the computer to learn from. To overcome this, the researchers used a generative artificial intelligence tool to create thousands of realistic, synthetic images of defects. These are not random drawings; they are high-fidelity simulations of cracks and bubbles that look and behave like real ones, filling the gaps in the training data. This allowed the system to learn from a vast library of examples, even though the physical lab only produced a small number of real defective pieces.

Once the system has identified a defect, it must decide how serious it is. This is where the concept of "grading" comes in. Instead of just saying "defect found," the system calculates a severity score based on the size, shape, and location of the flaw. A long, thin crack is treated differently from a small, round bubble. The system uses a dynamic threshold, meaning it adjusts its standards based on the specific batch of ceramics being inspected. If the raw materials change or the firing conditions shift slightly, the system recalibrates its definition of what is acceptable, ensuring that the grading remains fair and consistent regardless of external changes. This dynamic approach proved far more accurate than using a fixed, unchanging rule, which often leads to mistakes when conditions vary.

The entire process happens in a fraction of a second. The system was tested on a conveyor belt moving at a pace of about 30 pieces per minute, a speed that is manageable for a university lab but would be impossible for a human inspector to maintain with high accuracy over a long shift. The researchers tested the system on 360 different pieces, including flat tiles, tableware, and even low-relief artistic works with raised patterns. The results were striking. The system detected defects with an accuracy of nearly 90 percent, far surpassing the performance of human inspectors, who missed about 12 percent of visible flaws. The automated system reduced the missed detection rate to just under 3 percent. Furthermore, it could classify the type of defect and assign a grade with over 91 percent accuracy, matching the consensus of three independent human experts.

The system also proved its value in the real world of the laboratory. It was integrated into a workflow where it not only inspected the pieces but also sent the data to a central computer. This data was used to generate reports showing exactly where defects occurred and how often. By analyzing this data, the researchers were able to spot a hidden pattern: the speed at which the kiln heats up was directly linked to the number of bubbles appearing in the final product. This insight allowed them to adjust the firing process, reducing the number of bubbles by nearly a third in subsequent batches. This closed loop of inspection and process improvement is the true power of the system; it does not just find the bad pieces, it helps prevent them from being made in the first place.

There are, of course, limits to what this 2D vision system can do. When the ceramic piece has a complex, three-dimensional shape with deep grooves or raised reliefs, the system sometimes struggles. Shadows cast by the raised parts can look like cracks to the camera, leading to a slightly lower accuracy for these artistic pieces. The researchers acknowledge this and suggest that future versions could use 3D scanning technology to see around the corners and into the shadows, bridging the gap for even the most complex shapes. For now, however, the system represents a major leap forward for university ceramics. It offers a way to handle the small, diverse, and high-quality batches typical of academic and artistic production with a level of precision and speed that was previously unattainable.

In the end, this work is about more than just catching cracks. It is about preserving the integrity of the craft while embracing the tools of the future. By automating the tedious and error-prone task of inspection, the system allows the human experts in the university lab to focus on what they do best: teaching, creating, and pushing the boundaries of what is possible with clay and fire. The technology does not replace the artist; it supports them, ensuring that the final product is as perfect as the vision that created it. The system stands as a testament to the power of combining specialized optics, advanced artificial intelligence, and a deep understanding of the material to solve a problem that has plagued potters for centuries.