A computer vision quality-control pipeline for high-speed assembly lines. It flags scratches, misalignments, missing parts, and wrong labels, and pushes pass or fail into the plant MES.
Manual inspection was inconsistent between shifts and too slow for the line speed. Subtle defects showed up later as rework and returns.
The pipeline inspects every unit. Optimized PyTorch models and an OpenCV preprocessing stack run on edge devices on the shop floor. Operators see live overlays and alerts without slowing throughput.
Industrial cameras stream frames from defined checkpoints along the line.
OpenCV crops the region of interest, normalizes lighting, and corrects perspective before the model sees the frame.
A PyTorch detector looks for scratches, gaps, misalignments, foreign objects, and missing components.
The edge device draws the overlay and returns a pass or fail plus a defect code. The MES only receives that event.
Models are trained on real production footage and tuned for glare, vibration, and part variation. Inference stays on the edge so camera frames do not have to leave the factory network.
When an operator overrides a decision, that correction is logged and can be used to retrain. New lines reuse the same pipeline after the cameras are calibrated and a local dataset is captured.
No. Inference runs on edge hardware next to the line. The MES receives pass, fail, and a defect code.
Yes. The model library is trained per product family, with a severity threshold the plant can set.
The override is logged. That feedback is the set used when the model is retrained.
Empowering AI-Driven Automation & Transformation. Tell us whether you need a custom web build, an AI model, or a product like Voice AI Copilot.
3672 S 16th Ave, Arizona, 85713
info@atwebtechnologies.com
(480) 743-4017
