Researchers at IIIT-Bangalore have developed a computer vision system that can automatically detect defects on production lines by comparing each product to a single perfect reference sample. | Photo credit: SPECIAL ARRANGEMENT
Imagine if a factory could check every product on its assembly line just by comparing it to one perfect photo without training, huge data sets and expensive artificial intelligence (AI).
A team at the International Institute of Information Technology Bangalore (IIIT-B) has created a computer vision tool that can pick up even hairline scratches, dents and minor misalignments using a single reference image. The system was also showcased at the Bengaluru Tech Summit 2025.
A team of Jyotsna Bapat, Sasirekha GVK and H. Sanjeev of Integrated MTech built it for factories that face one recurring problem: inconsistent human control and costly AI-based inspection.
Workers are tired, lighting changes during shifts, and small factories don’t have the thousands of labeled images needed to train modern AI models. Most can’t even afford GPUs or specialized cameras.
This tool starts with a “golden reference image” – a high-quality photo of what the perfect product looks like. Every new product in the line is compared to this one image. Before comparison, the system adjusts each new product to exactly match the reference, using a technique called ECC alignment. Simply put, it “aligns” the new photo with the perfect one until every pixel clicks into place. It can automatically correct small rotations or tilts, so factories don’t need precision jigs or robotic arms.
The researchers say the idea came from watching small and medium-sized industries struggle with quality controls. Even when companies install AI systems, they become expensive and time-consuming to maintain because models need to be retrained whenever the product design changes or new defects appear.
The IIIT-B team wanted something simple—a tool that workers could understand, run on a basic computer, and not break when the lighting changed. Their biggest challenges included stabilizing ECC alignment for noisy textures, designing an anti-noise mask that works on reflective surfaces, and making sure the final output is easy to interpret for factory workers.
During alignment, the system also learns the natural behavior of the camera, its grain, small vibrations, sensor noise, and what reflections look like on that surface. This becomes the basic noise mask that tells the software what is normal for the camera and what is a real glitch. This is crucial because cheap cameras often produce false alarms when the light shifts slightly or when metal surfaces reflect differently. The basic mask filters out such problems.
Once the alignment is done, the system increases the brightness using CLAHE (a method that equalizes illumination), subtracts one image from the other pixel by pixel, checks how structurally similar they are, and then highlights only the differences that matter. These differences are displayed as a color-coded fault map, so even non-technical factory personnel can easily understand where the fault is.
Thanks to this careful filtering, the tool can pick up defects as small as a few pixels, even those that the human eye might miss. Works on flat parts, slightly curved surfaces, metals, reflective or semi-reflective materials and parts that move slightly on conveyor belts.
Factories can detect multiple types of defects at once, including scratches, dents, foreign particles, texture irregularities, shape mismatches, or rotational errors.
The tool was tested on different product categories under different lighting conditions. According to the team, it consistently achieves up to 98% accuracy with low false alarms and processes each image in less than 13 seconds on a regular CPU. No GPU needed and no AI retraining ever needed. The stringency can be adjusted depending on how sensitive the factory requires detection, helping to avoid unnecessary rejection of good items.
They believe this tool can help the industry reduce inspection times, reduce waste and make quality control more predictable. It also brings advanced automation within reach for small and medium-sized industries that often cannot afford deep learning-based inspection systems.
Published – 29 Nov 2025 21:50 IST
