Computer vision

Every detail matters.

Industrial quality control

Automated visual inspection that brings consistent defect detection to the production line.

YF / 01 Computer vision Industrial quality control — conceptual system illustration
System concept Illustration / 01
Industry
Automotive manufacturing
Our role
Vision system & line integration
Core tools
Industrial cameras · OpenCV · TensorRT
01 / The challenge

A practical problem.

Manual inspection varied between shifts, while a previous rule-based system rejected too many good parts. The new system needed to detect small defects on reflective aluminium without slowing production.

Engagement
5 months · Jan–May 2024
Project setting
Automotive supplier · West Midlands
02 / Our approach

Built around the constraints.

  1. Control the image.

    Synchronised cameras and diffused lighting give the model a consistent view of each part.

  2. Detect the detail.

    Custom vision models identify surface defects across the different part types.

  3. Close the loop.

    A GPU inference pipeline connects inspection results to the line’s reject mechanism and monitoring systems.

03 / The system

How it comes together.

Components → processing → output

A camera-to-production pipeline that turns consistent images into inspection decisions and traceable results.

  1. Capture

    Six cameras · Diffused LED lighting

    Synchronised views cover each part’s surfaces. Dome lighting reduces reflections on machined aluminium.

    OutputMulti-angle part images

  2. Prepare

    Noise reduction · Normalisation

    Images are cleaned and normalised before the model examines the part’s surface features.

    OutputConsistent inspection inputs

  3. Inspect

    EfficientNet-B3 · TensorRT

    Custom defect models identify and locate faults. Confidence thresholds distinguish acceptable parts from rejects.

    OutputDefect type, location & score

  4. Act

    PLC integration · Quality monitoring

    Inspection decisions trigger the reject mechanism, while per-part logs support traceability and shift reporting.

    OutputAccept/reject signal & record

04 / The outcome

Results from this project.

Defect recall
99.4%Compared with 87% for manual inspection
Processing per part
<60 msEnd-to-end inspection pipeline
False rejection rate
0.6%Compared with 12% for the rule-based system

Before

Manual checks varied between shifts; rule-based inspection rejected too many good parts.

After

Consistent automated inspection, with per-part traceability and shift-level quality reporting.

In context. Reflective, chrome-plated parts reduced detection accuracy to 96.2%. Consistent lighting and calibration remain important to the reported results.

Your next project

Working on something similar?

Tell us about your system and where you need support.

Discuss your project