Overview AI's OV Defect Creator Studio matches real defect photos in PCB benchmark

Overview AI has published benchmark results testing whether defect images generated by its OV Defect Creator Studio tool can substitute for real defect photos when training a printed circuit board inspection model. Using the public VisA PCB1 dataset, the company found that adding 50 generated defect images to a pool of just 10 real ones let a segmentation model match, and in one run exceed, the results produced by 50 real defect photos alone.

The test used a stock, open-source U-Net segmentation model with a ResNet-34 encoder and 512 x 512 inputs, scored against a 2 percent false-reject limit so a model couldn't inflate its score by flagging more good boards as bad. Trained on 50 real defect photos covering scratches, bent pins, missing components, and melt defects, the model averaged an 84 percent catch rate. With only 10 real photos, the average fell to 58 percent. Adding 50 generated images to those same 10 real photos brought one training run back to 84 percent and another to 86 percent, using 40 fewer real defect photos than the original benchmark. For a defect library that can take months to fill with real scrap, that gap can decide whether an inspection line gets validated in weeks or quarters.

The gains varied by defect type. Scratch detection, tested on 10 real photos, rose from 33 percent to 75 percent once the 50 generated images were added. Missing-component detection climbed from 31 percent to 58 percent, bent-pin detection from 76 percent to 96 percent, and melt detection from 69 percent to 84 percent.

Overview AI noted that the benchmark measures the training data itself, using a fixed model, rather than its own camera hardware or inspection software. It also found a gap in its synthetic melt defects: real examples show deposits around pins and pads and on the metal crystal can, while the generated versions concentrated on pins and pads alone, a specific target for the next round of generation.

Before training, Overview AI standardized board orientation across the dataset, since PCB1's boards appear in two opposing orientations that otherwise correlate with the good or defective label. The company framed the results as a workflow: start with whatever real defect photos exist, use OV Defect Creator Studio to generate examples that fill in underrepresented categories, then validate the trained model against real production images before it reaches the line.