Skip to content
Deploint

Case study 05 / Manufacturing

Computer Vision Quality Inspection

Line-side cameras and edge inference that flag surface defects in real time, with operator feedback loops that continuously improve the model.

Concept Architecture

Illustrative reference architecture, not a client engagement. Outcomes described here are design goals, not measured results.

Talk to an EngineerExplore the architecture

Reference flow

05 stages

Reference flow for Computer Vision Quality Inspection: Cameras, then Edge inference, then Defect triage, then Operator feedback, then Retraining.

01Challenge

Why this problem is hard.

Manual visual inspection of surface defects is tiring, varies between shifts and is hard to scale with line speed. Rule-based machine vision handles well-defined checks but struggles with variable defects such as scratches, dents, contamination and texture anomalies. The plant needs consistent inspection at line speed without stopping production for every uncertain call.

  • 01

    Line-speed decisions

    Each part must be classified within the takt time, before it reaches the next station or the reject gate.

  • 02

    Rare, variable defects

    Some defect types appear infrequently, which makes collecting training examples slow.

  • 03

    Changing conditions

    Lighting, part variants, supplier changes and lens contamination shift the image distribution over time.

  • 04

    Traceability

    Every decision must link to the part serial or batch for the quality record.

02Business context

Assumptions and constraints.

The concept assumes a discrete manufacturing line with an existing MES, a PLC-controlled reject mechanism and quality engineers who own defect definitions. Inspection augments the quality process; quality engineering keeps authority over criteria and dispositions.

Constraints the design must respect

01Deterministic timing
Inference and the reject signal complete inside the line's timing window, every time.
02Plant-floor resilience
The station keeps working when the plant network or cloud connection is unavailable.
03Human authority
Uncertain predictions go to an operator. Quality engineers approve every model before it reaches the line.
04Change control
Model updates follow the same change control as other process changes, with rollback.

03Architecture

Reference architecture.

Seven stages from camera trigger to a retrained model. The system has three outcomes, pass, reject and review, so the model never has to pretend to be certain. Select a stage to see what it does and the components involved.

Industrial cameras with controlled lighting capture each part on a hardware trigger. Optics and lighting are engineered first, because they determine what any model can see.

  • Industrial cameras
  • Structured lighting
  • Hardware triggering
  • Enclosures

Key patterns

  • Pass / reject / review triage
  • Edge inference
  • Human-in-the-loop labeling
  • Models under change control

04Technology

Representative technology.

Technologies we would consider for this concept. Final selection follows the organization's existing standards, skills and estate. Listing a technology does not imply a vendor relationship.

  • 01

    04 items

    Vision hardware

    • Area-scan or line-scan cameras
    • Controlled lighting
    • Hardware triggers
    • Edge accelerators

  • 02

    04 items

    Models

    • Detection and segmentation models
    • Anomaly detection for unseen defects
    • ONNX export
    • Quantization for edge runtimes

  • 03

    04 items

    Integration

    • PLC I/O
    • OPC UA
    • MES connectors
    • Message queue for results

  • 04

    04 items

    MLOps

    • Labeling tool
    • Dataset versioning
    • Model registry
    • Fleet deployment and rollback

05Engineering approach

Principles behind the design.

The decisions that shape every stage of the architecture, and why we would hold to them under delivery pressure.
  1. 01

    Optics before algorithms

    Most inspection problems are won or lost at the camera and lighting stage, before any training.

  2. 02

    Design for uncertainty

    Three outcomes, pass, reject and review, keep uncertain calls with people instead of forcing a guess.

  3. 03

    Operators in the loop

    Line-side review is designed for speed and ergonomics, because operator labels are the data that improves the model.

  4. 04

    Models under change control

    Every model version is benchmarked, approved and reversible, like any other process change.

06Implementation

A phased delivery path.

Each phase ends with a working, reviewable increment and an explicit decision on whether and how to continue.

Delivery path

04 phases

Delivery phases: Feasibility, then Pilot station, then Closed loop, then Fleet rollout
  1. 01Feasibility
  2. 02Pilot station
  3. 03Closed loop
  4. 04Fleet rollout
  1. Phase 0101

    Feasibility

    Image real parts and defects under candidate lighting, agree defect definitions and test whether defects are visible and separable.

    Deliverables

    • Imaging study
    • Defect taxonomy
    • Feasibility report
    • Station design
  2. Phase 0202

    Pilot station

    Install one station in shadow mode: it classifies but does not reject, and results are compared with manual inspection.

    Deliverables

    • Pilot station
    • Baseline model
    • Shadow comparison
    • Review interface
  3. Phase 0303

    Closed loop

    Enable automatic reject for well-understood defects, route uncertain cases to review and start the retraining cycle.

    Deliverables

    • PLC integration
    • MES write-back
    • Retraining pipeline
    • Change control procedure
  4. Phase 0404

    Fleet rollout

    Replicate to further lines with central model management, monitoring and per-station calibration.

    Deliverables

    • Station template
    • Fleet management
    • Drift monitoring
    • Support model

07Intended outcomes

Design goals, not results.

What this architecture is designed to achieve. These are intended outcomes for a concept, not measured results from a client engagement.

  • Design goal 01Intended

    Consistent criteria

    Every part is assessed against the same approved defect definitions across shifts and lines.

  • Design goal 02Intended

    Inspection at line speed

    Routine calls are made inside the line's timing window without stopping production.

  • Design goal 03Intended

    Focused human attention

    Operators spend their time on uncertain cases, where their judgment adds the most.

  • Design goal 04Intended

    A model that keeps up

    Operator feedback and drift monitoring keep the model aligned with changes in parts, suppliers and conditions.

Concept Architecture

No figures are attached to these goals. Actual results depend on the environment, data and delivery, and would be measured against baselines agreed at the start of a real engagement.

08Lessons & risks

Engineering lessons and risks to manage.

Where designs like this tend to run into trouble, and what we would plan for from the start.

  • Risk 01

    Definitions need agreement

    If inspectors disagree on what counts as a defect, the labels will too. Align definitions before training.

  • Risk 02

    Rare defects need a strategy

    Anomaly detection, augmentation and targeted collection help when examples are scarce.

  • Risk 03

    Edge fleets need operations

    Many stations mean patching, monitoring and rollback at fleet scale. Plan for it from the pilot.

  • Risk 04

    Shadow mode builds trust

    Comparing the system with manual inspection before it acts gives quality teams evidence on their own terms.

Start an engineering conversation

Facing a similar engineering problem?

Bring us the problem and the constraints around it. We'll help architect the system.

Talk to an EngineerView Case Studies