Computer vision proof of concept

A short feasibility study testing whether an overhead camera and a defect-detection model could catch the surface flaws a manual inspection line was missing.

Timeline: June 2024 - September 2024
Detection accuracy: 94% on test set
Line slowdown: None

An overhead camera rig positioned above a manufacturing line. — Image adapted from the CIDS project image set for the project detail page.

The proof of concept let us see the failure cases before we committed budget, not after.

QL

Quality lead

Manufacturing partner

Recognition and proof

Project signals

Sensor readings displayed on a tablet beside plant machinery.

Timeline

Feasibility confirmed within one sprint

A working detection model on the partner's own sample images was demonstrated within four weeks of project start, ahead of the original estimate.

Researchers working through diagrams during a funded discovery project.

Outcome

Informed a production investment case

Results from the proof of concept became the evidence base for the partner's board paper requesting funding for a full line deployment.

Project overview

Testing feasibility before committing to a full deployment.

The partner’s manual inspection line was missing a class of small surface defects that only showed up under specific lighting angles, and wanted to know whether computer vision could catch them before committing to hardware.

CIDS built a small labelled dataset from the partner’s own reject bin, trained a detection model, and ran it against a held-out test set alongside the existing manual process to compare catch rates directly.

The model matched or exceeded human inspectors on the defect classes in scope, which was enough evidence for the partner to greenlight a follow-on project to integrate a camera rig into the live line.

Industry partners reviewing inspection results together. — The proof of concept ran against archived reject-bin images before any camera hardware was installed on the live line, so results could be judged without needing production access.94% detection accuracy on held-out test set

Before and after

2 operational shifts from the project

Computer Vision Proof of Concept

Before

Investment decision

No evidence base for a line-wide camera rollout.

Defect detection

Manual inspectors, inconsistent under certain lighting.

After

Investment decision

A validated accuracy figure and defect catalogue.

The partner could size the investment before committing to hardware.

Defect detection

Model matched or exceeded manual catch rate on the flagged classes.

The gap in current inspection is quantified, not just suspected.