Manufacturing & auto components
Surface-defect detection on the production line
Camera-based quality inspection that finds and classifies surface defects at pixel precision, at line speed.
This page describes a problem class and the architecture we deploy for it, written against a named public benchmark — the open dataset or competition where the world's data scientists have tested approaches to this exact problem against a hard metric. It is not a client engagement, and the benchmark's results, prizes and rankings belong to its host and participants, not to us. Client work is confidential and is only ever published with written permission.
The problem
What it costs when this goes unsolved.
Human visual inspection puts a ceiling on line throughput and still misses defects at the end of a shift. A missed surface defect ships to a customer; an over-called one scraps good material. Both have a unit cost that compounds across every metre of output, and neither improves by adding inspectors.
Production camera imagery is heavily imbalanced — most frames show clean material, and the defects that matter may appear in a fraction of a percent of them. Labels are masks drawn by process engineers with genuine disagreement at defect boundaries. A model trained without gating on the clean majority learns to hallucinate defects; one validated on the wrong split memorises the camera, not the metallurgy.
How we build it
The architecture, stage by stage.
01
Gate before you segment
A fast classifier screens out the clean majority of frames so the expensive segmentation model only runs where something might be wrong — the two-stage cascade that production vision QC settled on.
02
Pixel-level segmentation
U-Net and FPN decoders over pretrained encoders produce a mask per defect class, not just a verdict — so the output is inspectable by the process engineer who has to trust it.
03
Thresholds set by cost, not accuracy
The operating point is chosen against your scrap cost and escape cost, per defect class, and revisited once real line data arrives.
04
Line-side deployment
Quantised inference on edge hardware beside the camera, with alerts into the line's existing andon or MES flow rather than another dashboard nobody watches.
Model families on this problem class: U-Net / FPN segmentation · Pretrained CNN encoders · Two-stage cascade · Edge-quantised inference.
What you get
What an engagement hands over.
Everything below goes in the scope document before you sign it, with a fixed price or a rate with a ceiling — the same terms as every other engagement in the catalogue.
- A defect-detection model with per-class masks and severity grading
- An operating-point analysis priced in your scrap and escape costs
- Edge deployment beside the camera, integrated with the line's alerting
- A labelling protocol your QC team can run to keep the model current
- Retraining cadence and drift monitoring, documented for your team
Provenance
The benchmark behind this page.
Severstal Steel Defect Detection, run by Severstal, is the public proving ground for this problem class. The figures below are the host's, cited as context for how seriously this problem is tested in the open — they are not our results and we do not claim them.
How to buy this
The services this build draws on.
More proof
Next step
Thirty minutes on whether this fits your problem.
Bring the constraint — the regulator, the data boundary, the latency budget. If your data cannot support this build yet, the call will conclude with what to fix first, not with a proposal.