DeepVision-Web: AI inspection on an unseen board
The AI inspection now spots defects well on a board it has never seen: AUC 0.96 on the top side, 0.88 on the bottom. On the top, averaged over two photos, it finds 80 % of the defects at 4.7 % false alarms; the bottom is weaker, mostly at through-hole solder joints and empty pads. That is not yet enough for a verdict of its own, so the AI still only reports a “suspicion”. Averaging several photos of the same board helps most (false alarms from 15 % with one photo to 6.5 % with five); online datasets, more training data or larger crops did not. Next: a third prepared board, a model for through-hole joints and measurements on the phone. Private project — only occasional previews.
aoiaievaluation