Quality Control Vision

This page explains how to design and deploy AI/Machine Vision for Quality Control that actually works on production lines—measurable, auditable, and maintainable (no hype).

What it is

Core Capabilities

Quality Control (QC) Vision uses cameras, optics, lighting, and algorithms to detect defects, verify assembly, read text/codes, and perform dimensional checks at line speed.

Methods Include

Classification - OK/NG, defect type
Detection - boxes around defects/parts
Segmentation - pixel-precise defect regions
Metrology - dimensions/angles/gaps; 2D/3D
Anomaly detection - learn "normal" then flag deviations

Where it fits

Surface Defects

Scratches, dents, contamination
Coating/paint issues

Assembly Verification

Presence/absence, orientation
Connector seats, torque marks

Packaging & Labeling

Print quality, OCR/lot/date codes
Barcode/QR, label skew/mismatch

Seals & Integrity

Cap/foil seals, blister pack
Film wrinkles/tears

Dimensional Checks

Gap/flush, holes/slots distance
Thread/teeth count

Process Cues

Fill level, color drift
Baking/curing state, solder/weld beads

Optics & lighting that actually work

Critical Parameters

Pixel Density

Size of the smallest feature ≥ 3–5 pixels across (detection), 8–12 px (reliable classification), 15–25 px (metrology/OCR).

Angles & Motion

Keep view angles ≤ 25–30°, sensor roll ≤ 5° for dimensional/OCR tasks. Freeze motion (1/250–1/2000 s); use encoders + strobe for conveyors.

Lighting Choices

Backlight: for silhouettes/edges/gaps

Coaxial/Brightfield: flat reflective surfaces/print

Dome/Diffuse: glossy/curved parts

Dark-field/Ring: highlight scratches/texture

Polarizers (crossed): suppress glare

Lenses & 3D Options

Use telecentric for accurate dimensional metrology; varifocal for general inspection. 3D options: laser triangulation/structured light for height/warp; ToF for coarse z.

Algorithms & modes

Classical Vision

Threshold/edges/template matching for stable, high-contrast tasks.

Deep Learning

CNN/transformers for variable textures, complex defects.

Hybrid Approach

Classical pre-processing + deep models + rule validators (tolerances/regex/check digits).

Metrology

Calibrated pixel-to-mm, sub-pixel edge fit; 3D height maps for warp/bow.

Data & ground truth

Requirements Planning

• Defect spec first: names, visuals, tolerances, cost of false reject vs false accept.
• Representative samples: capture day/night, lot-to-lot, tool wear, color batches.
• Sample size: start 200–500 images per defect state; grow with failure analysis.

Data Management

• Annotate consistently: pixel masks for surface defects; line/point sets for metrology.
• Splits by site/time/device to avoid leakage; hold out a golden set for acceptance.

Metrics that matter

Detection/Segmentation

Per-defect Precision/Recall/F1
mAP, PR curves
False rejects/accepts per 10k

Metrology

Bias, repeatability, reproducibility (Gage R&R)
Uncertainty (mm) vs tolerance
Cp/Cpk impact

Operations

End-to-end latency
Throughput (parts/min)
Image retention %, downtime due to vision

Throughput & timing

Full Path Budget

Budget the full path: exposure → transfer → preprocess → inference → postprocess → PLC I/O → rejector actuation.

Strobe Optimization

Use strobe to cut exposure while keeping brightness.

Part Tracking

Buffer & track parts so ejectors hit the right one (distance → time).

Timing Budget

Keep per-part latency < cycle time with headroom (e.g., 60–70%).

Integration & traceability

I/O & Protocols

PLC EtherNet/IP, PROFINET, Modbus/TCP

OPC UA Industrial communication standard

MQTT Event messaging

Reject Stations

Air-blast/diverter/marker; interlock with safety systems.

Evidence & Data Management

•
Evidence: store crops (defect region) + minimal context; configurable retention (e.g., 30–90 days); hash for integrity.
•
SPC & MES: push pass/fail and measurements; enable trend alarms and auto-stop rules.

Reliability & upkeep

Preventive

Lens clean, focus checks
Strobe aging, fan/temperature alarms

Re-qualification

Periodic MSA/Gage R&R
Test chart runs, lighting re-tune

Drift Control

Monitor confidence and class distributions
Trigger retraining with active learning

Spares

Spare lights/cameras/lenses
Printed SOP for swap-and-calibrate

Validation roadmap

Spec & risk

defect list, tolerances, FA/FR costs

Optics pilot

lighting/lens proof with target px/feature

Model baseline

train/val/test; report per-defect F1 & uncertainty

MSA/Line test

Gage R&R, latency budget, eject accuracy

Acceptance

hit KPIs on golden set + 1–2 weeks live shadow run

Handover

SOPs, retrain triggers, spare kit, dashboards

Red flags

"Works in any lighting/angle" without a pixel-density or lighting plan

Model-only FPS (no decode/IO/post-proc/PLC)

No per-defect metrics; only overall accuracy

No MSA/Gage R&R; no golden-set acceptance

One camera to do overview + detail + OCR simultaneously

GaugeSnap integration

Edge AI Packs

Surface-defect segmentation
OCR/lot, label/print check
Cap/foil seal validation
Gap/flush metrology (2D/3D)

Industrial Connectors

PLC (EtherNet/IP, PROFINET, Modbus/TCP)
OPC UA, MQTT → MES/ERP/SPC

Dashboards

Per-defect F1, mAP/PR curves
Latency histograms, eject accuracy
Drift alerts

Sustainable AI

INT8/FP16, ROI pipelines
kWh/1k inferences and cost KPIs

Example event

{
  "event": "qc_result",
  "station_id": "lineB_cam2",
  "part_sn": "ABX-24-008173",
  "result": "NG",
  "defects": [
    {"type":"scratch","mask_bbox":[412,96,120,64],"score":0.91},
    {"type":"label_skew","angle_deg":5.7,"score":0.88}
  ],
  "measurements": {"gap_mm": 0.42, "flush_mm": 0.10},
  "latency_ms": 62,
  "conveyor_speed_mps": 0.8,
  "ts": "2025-08-25T12:34:56Z"
}

How to start (low-risk)

Send us:

Defect Spec Sheet

Names, tolerances, costs of FR/FA

Videos & Photos

2–3 min videos per station (day/night, speeds) + photos of parts

Line Data

Cycle time, conveyor speed, available PLC I/O

We'll return:

• Optics/lighting plan
• Pixel-density check
• Baseline metrics report (per-defect F1, latency)
• Pilot with clear KPIs

Principle

Prove with optics + per-defect metrics on your line—then scale.