Компьютерное зрение

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34import cv2import numpy as npfrom inspection import Detector, Verdict, plc# Loaded once at start-up: model load is ~900 ms and the line gives us 240 ms# per part. Anything that can be done before the first part arrives, is.detector = Detector("models/weld-seam-v7.onnx", provider="CUDAExecutionProvider")def inspect(frame: np.ndarray) -> Verdict:    # Crop to the fixture. The camera sees the whole cell, but the weld is always    # in the same window, and cropping first cuts inference cost roughly fourfold.    roi = frame[220:820, 340:1240]    # Normalise for the lighting, which drifts all shift as the lamps warm up.    # Without this the model quietly gets worse between 06:00 and 10:00.    gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)    gray = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8)).apply(gray)    # Infer. "Not sure" is a distinct outcome from "no defect", and conflating    # the two is how vision systems lose the operators' trust in week three.    boxes, scores = detector.run(gray)    confident = scores >= 0.62    if not confident.any() and scores.max(initial=0) > 0.35:        return Verdict.REVIEW    # Decide inside the cycle time. The PLC holds the authority here; the vision    # system only ever reports, so a crash fails safe instead of passing scrap.    if boxes[confident].size:        plc.write("Line1.Reject", True)        return Verdict.REJECT    plc.write("Line1.Pass", True)    return Verdict.PASS