DeepVision-Web: an AOI autopilot that tests and tunes itself
New plan: instead of me checking photos by hand, an autopilot runs phone photos, videos and scans through the real app, checks whether each component box really sits on its part, rates the crops and then tunes thresholds, references and the dataset itself. Only a few labels from the phone are still up to me. Private project — only occasional previews.
Where it started: on slightly tilted phone photos the component boxes often did not sit exactly on the parts. One pose for the whole board cannot follow the lens, the angle or a bent board, so a crop confirmed as OK sometimes showed bare board instead of the component. That gets a local correction field, and the fit of every crop is measured against a reference scan.
From there it grows into a loop: run the images, measure the fit (green, yellow or red per image), rate the crops, tune, and check again. A new database of top and bottom photos with good parts and real defects helps, and for the first time it gives the bottom side a dataset of its own.
Nothing is adopted blindly. Every change has to pass a regression first: simulated phone photos with tilt, lens distortion, a bent board, blur and changing light, where the true position of every part is known; real photos; a set of confirmed defects and false alarms; and the labels from the phone. Only then does the autopilot write new thresholds and references, and I get the report afterwards.
What counts as a defect: a missing or wrong part, an offset or rotated one, solder defects and foreign objects. A part that sits off by a quarter of its pad width or more counts as offset. A run takes up to about 15 minutes and starts when I ask for it. The plan is written; the building comes next, step by step.