Eine anlernbare Kamerainspektion in vier Schritten: Quelle wählen, Bereich markieren, ein paar gute und schlechte Teile zeigen, prüfen lassen. Webcam oder synthetisches Förderband, alles bleibt auf Ihrem Gerät. Achten Sie darauf, wie wenige Beispiele reichen und wo die Konfidenz unsicher wird.
A trainable camera inspection in four steps: choose a source, mark a region, show a few good and bad parts, let it inspect. Webcam or synthetic conveyor, everything stays on your device. Note how few examples suffice and where confidence becomes uncertain.
Show the system a few good and a few bad parts, and it starts sorting OK from NOK — the classic shop-floor workflow of a trainable camera inspector, reduced to its essence. Use your webcam or the built-in synthetic conveyor.
The region of interest is resampled to 64 × 64 pixels. From it, three feature blocks are computed: a hue–saturation histogram (12 × 4 bins),
a brightness histogram (8 bins) and a coarse histogram of gradient orientations (4 × 4 cells × 6 directions, HOG-style). Each block is L2-normalised,
the concatenated vector is the part's fingerprint. Classification is a weighted k = 3 nearest-neighbour vote; confidence combines the vote share
with how much closer the winning class is than the losing one. The desktop prototype uses the same feature idea with a random forest (scikit-learn) and OpenCV.
Privacy: the webcam stream is processed frame by frame inside this page. No frame, sample or result is transmitted or stored anywhere; closing the tab erases everything.