Limitations

• What both models miss is the same thing: low-contrast contamination lying flat on the base, the hardest case for rebuild-and-compare scoring. A real line would add a backlight.
• False alarms are cheap by design: a flagged good bottle costs an operator a few seconds of review; a missed defect costs a lot more. Both alarm levels are tuned accordingly.
• No cheating: defects never touched training or the alarm levels, and the test set was graded exactly once, for both models.
• Simpler ideas we tried and dropped: scoring bottles by their distance to the most similar good bottles (nearest neighbours) caught only 12 of 63 defects; rebuilding beats lookup here.
• Verdict: both are solid screening alarms (every break caught by the neural net, and each alarm comes with a picture of why), but neither is a safety guarantee on its own.

Built in a week, by five people

This project was our entry to the Imperial College London AI & Autonomous Industrial Systems Summer Programme, a week-long project sprint in teams of five, judged across ten teams. Our entry took first place.

The two contenders on this site began as independent tracks within the team: the PCA pattern-matcher and the patch autoencoder were built separately, then brought together on the same data split and the same rules for the head-to-head. Every number and figure on this site is the actual output of the two project notebooks.