OV Defect Creator Studio · VisA PCB1 benchmark

Can you train a defect detection model with synthetic data?

Overview AI·September 25, 2026

We benchmarked it. Ten real defect photos plus 50 images generated in OV Defect Creator Studio matched the 50-real-photo benchmark at 84% on one seed. On another, the model outperformed it, catching 86% of defects.

A fifth of the real defect data.

The bottleneck: collecting real defect images

A visual inspection project can have a camera, lighting, and a clear definition of a bad part, yet still wait on training data. Good parts are easy to photograph. A particular scratch, bent pin, or missing component may take weeks or months to appear often enough to build a useful defect library.

Teams either wait for scrap or damage parts on purpose. Synthetic defect generation offers another way to collect training examples. The question we wanted to answer was concrete: can those generated images help a model catch defects in real photographs?

What we benchmarked

We used PCB1 from the public VisA dataset, comparing real defect photos with images generated in OV Defect Creator Studio. The training pool contained 50 real defect images and 50 generated images, covering scratches, bent pins, missing components, and melt defects.

A false reject is a good part incorrectly flagged as defective. Holding that limit fixed makes the comparison useful: a model cannot earn a better result simply by rejecting more good parts. The experiment measures the value of the training data using a common model.

The result: matching and exceeding the real-data benchmark

Training with 50 real defect photos produced an average catch rate of 84%. With only 10 real defect photos, the average was 58%. That gap is the problem a team faces when its real defect collection is still small.

Adding 50 generated images to those 10 real defect photos changed the result. Across multiple seeds, one run matched the 84% benchmark. Another reached 86%, exceeding it with 40 fewer real defect photos.

These individual runs show that generated examples can help a detector reach the real-data benchmark with a much smaller collection of real defects. For manufacturers, that means a useful route to expanding training coverage while rare examples are still difficult to obtain.

Scratch detection more than doubled

The clearest improvement was in scratches. With 10 real defect images, the model caught 33% of scratch defects. Adding 50 generated images raised that to 75%, without collecting a single additional real defect photo.

Missing-component detection also rose from 31% to 58%. Bent-pin detection improved from 76% to 96%, and melt detection from 69% to 84%. All four defect categories improved in this 10-real-image comparison.

This is where synthetic data becomes especially useful: it gives the model more examples of defects that a small real dataset does not represent well. Instead of waiting for another damaged board, an engineer can generate targeted variations and measure whether they help.

Look at the real and generated defects

The examples below come from the study itself. Choose a defect type to compare real VisA photographs with generated training examples. The outlines are annotations identifying the defect regions, not predictions from the detector.

ScratchesBent pinsMissing componentsMelt defects

Scratches: real vs. generated

33% → 75% caught

Swipe across the examples, or open the full-size image below.

Look for pale abrasions where the blue coating has been scraped away. The examples vary in length, direction, and the number of parallel marks. Teaching this range is more useful than repeating one obvious gouge.

Open full-size scratches examples ↗

Catch rates are rounded averages across seeds for 10 real defect images before and after adding 50 generated images, at a 2% false-reject limit. Images derived from VisA, Zou et al., under CC BY 4.0. Cropping, annotations, and synthetic modifications by Overview AI.

How we built the generated training set

Each generated image began with a defect-free board from the training side of the split. The operator used OV Defect Creator Studio to place a defect on that board, then painted a segmentation mask around the defective material. The generated set covered the same four categories as the real training pool, including a few boards with more than one defect.

The visual details mattered. On this PCB, a scratch is typically a pale abrasion through the coating. A bent pin involves displacement of the leg. Missing components expose surfaces or pads that were previously covered. Those are specific appearances a training set needs to represent, beyond the broad labels “scratch” or “bent.”

The melt examples show why category coverage alone is not enough. Real examples include deposits around pins and pads as well as on the metal crystal can, while the generated examples shown above are concentrated around the pins and pads. Inspecting that difference gives the engineer a concrete direction for the next round of generation: add the missing appearance, not just more copies of the same one.

What the numbers mean for visual inspection

The result supports a practical workflow: start with the real defect images you have, use generated examples to fill gaps, and validate on real images. A model missing scratches needs a better scratch library. A model missing a particular bent-pin shape needs examples of that shape.

This benchmark used a stock U-Net on one PCB dataset. Its percentages describe that experiment, rather than the performance of an Overview AI camera or a complete production inspection system. Keeping the model fixed lets us examine what the generated training data contributes.

The opportunity is to shorten the wait for useful training examples. Teams can build coverage for known rare failure modes earlier, then keep refining the inspection as more real production images become available.

Frequently asked questions