CASE STUDY · COMPUTER VISION APPLICATIONS
A computer vision quality-inspection line for a manufacturing client.
- 82%Defect detection up from 82% to 94%
- 4%False rejects below 4%
01 · PROJECT OVERVIEW
The client and the brief.
An anonymised manufacturer. Name withheld under confidentiality.
Existing line cameras, a defect-detection model and PLC-linked alerts. The system inspects each part as it passes, flags suspect items and signals the line controller to divert them. Operators see the flagged image and make the final call on borderline cases.
- 1Visual checks missed defects.
- 2Inspectors stood at the end of a fast line and judged parts by eye.
- 3Detection varied between shifts and between people, and it dropped late in a shift.
- 4Some defects were only found after parts reached the customer.
- 5The plant wanted a more consistent check without slowing the line or filling it with false rejects, which waste good product.
02 · WHAT WENT WRONG
What went wrong, and how we fixed it.
Every engagement hits problems. These were the ones that mattered here, and what we changed so they did not come back.
- PROBLEM 01In shadow mode, false alarms spiked in the afternoon
The cause was sunlight through a side door adding glare to the surface of the parts. The model was fine. The data was not representative.
- PROBLEM 02We added shading over the inspection point, collected more footage in bright conditions, relabelled and retrained
We also added a monitoring check on image brightness, so the system warns when lighting drifts outside the range it was trained on.
- PROBLEM 03A second issue came up after launch
A product variant with a glossy finish produced reflections the model had never seen, and it flagged good parts. Because every decision was logged with its image, quality staff spotted the pattern within a day. We collected examples of the new finish, labelled them and retrained.
- PROBLEM 04We also changed the process so that a new variant is run in shadow mode for a few days before it is allowed to reject anything
03 · THE PROCESS
How we ran it.
- 01Weeks 1 to 2
site visit, review of camera position and lighting, and a session with quality staff to agree what counts as a defect. We wrote it down with example images.
- 02Weeks 3 to 6
captured footage across shifts and batches, labelled it with the plant's inspectors reviewing a sample of the labels, and trained a first model.
- 03Weeks 7 to 9
shadow mode on the live line. The model made calls but did not reject anything. We compared it with inspectors every day.
- 04Weeks 10 to 12
connected the reject mechanism, set thresholds and trained operators.
- 05Step 5
Weekly review meetings with quality and maintenance throughout.
04 · RESULTS
What changed.
Measured against the way the work ran before the engagement began.
- NOW 01
Quality staff review the flagged images each morning and relabel disputed cases, which feeds the next training round.
- NOW 02
Inspectors spend their time on borderline parts and root-cause work rather than staring at the line.
- NOW 03
Changes to product or lighting trigger a model check before production, not after problems appear.
- NOW 04
Defect trends by shift and batch are visible to production managers, who use them in supplier and process discussions.
- 82%
Defect detection up from 82% to 94%
- 4%
False rejects below 4%
05 · WHAT WE BUILT
How it works.
- Existing cameras were kept, with a fixed shield and a stable light source added over the inspection point, because lighting was the biggest source of variation.
- An object-detection and classification model trained on labelled examples of good parts and each defect type.
- Inference on an industrial edge computer next to the line, so decisions were made in time and no video left the site.
- An interface to the PLC for reject signals and alerts, with a fail-safe: if the system stopped, the line reverted to manual checks and raised a warning.
- Logging of every decision with image, score and outcome, so quality staff could audit and relabel.
- Two thresholds: one for automatic reject and one for "send to human review".
- Monitoring and handover
- We trained two engineers from the plant's maintenance team to restart the service, replace a camera and check the lighting. A one-page runbook covers what to do when the system raises a warning. We set up a simple dashboard of rejects by defect type, brightness drift and the share of low-confidence calls, and agreed a retraining review each quarter.
06 · QUICK ANSWERS
The short version.
- What problem did the client have?
- Visual checks missed defects. Inspectors stood at the end of a fast line and judged parts by eye. Detection varied between shifts and between people, and it dropped late in a shift. Some defects were only found after parts reached the customer. The plant wanted a more consistent check without slowing the line or filling it with false rejects, which waste good product.
- What did Brandsmashers build?
- Existing line cameras, a defect-detection model and PLC-linked alerts. The system inspects each part as it passes, flags suspect items and signals the line controller to divert them. Operators see the flagged image and make the final call on borderline cases.
- What were the results?
- Defect detection up from 82% to 94%; false rejects below 4%.
YOUR TURN
Have a similar requirement?
Tell us what is stuck. We will come back with a plan and matched, pre-vetted engineers.
