Anomaly DetectionWeld InspectionSurface InspectionVideo QC/QAAI Assistant
HOME/PRODUCTS/LINE SURFACE
LINE SURFACE · Surface Defect Detection

Trained only on normal —
it still catches defects it has never seen

LINE SURFACE learns the distribution of normal product images and exposes anything different as a pixel-level heatmap. Start with 30 defect samples — or none — and stay protected against novel defect types.

99.2%
verdict accuracy
1.8%
overkill rate
4 weeks
time to deployment
INPUT

Imaging & capture

Process-tailored lighting and optics maximize defect contrast

LEARN

Normal-distribution learning

Patch-level embedding distribution modeled from normal images

HEATMAP

Anomaly heatmap & score

Where and how much it differs, at pixel level

JUDGE

Verdict · grading · report

Pass/fail + defect grade + MES history

/01 How it works

It does not memorize defects. It understands normal.

Collecting thousands of images per defect type collapses the moment a new defect appears. LINE SURFACE goes the other way — learn the distribution of normal, and surface everything outside it as a defect candidate.

01 · IMAGE

Optimize imaging

Lighting angle, wavelength and optics designed per defect family to secure contrast first

02 · EMBED

Learn normal embeddings

Patch-level normal distribution modeled from a few hundred normal images — no defect labels

03 · MAP

Heatmap & score

Distance from the normal distribution rendered as a pixel heatmap — the verdict evidence is visible

04 · GRADE

Verdict & grading

Heatmap-based pass/fail with automatic defect size, location and severity records

* Real defects collected in operation are fed back into validation sets and the type classifier.

ONE-CLASS

Normal-only training

200–1,000 normal images are enough to start. Never delay a launch waiting for defect data — ideal for new lines and new products.

HEATMAP

Evidence you can see

The heatmap shows why it is NG. One visual artifact serves inspector verification, process root-cause work, and customer claim evidence.

NOVEL

Novel-defect coverage

Unseen defect types are still flagged as "different from normal" — breaking the cycle of rebuilding models after mass escapes.

TUNING

Overkill management

The line between acceptable blemish and true defect differs per plant. Region-wise sensitivity maps and threshold tuning keep overkill controlled.

/02 Use Cases

Use Case Analysis

Coated webs, molded exteriors, continuous sheets — anything with a surface qualifies.

Batteries · Electrode coating

Started with 30 defect samples, on the line in 4 weeks

Challenge

A new electrode line had only 30 defect samples; "we need thousands" answers had stalled adoption entirely.

Approach

The anomaly model was trained on normal coating images only; the 30 defects were used purely for threshold validation.

Result

99.2% accuracy at 1.8% overkill, deployed in four weeks. Never-collected novel defects were caught before first escape.

Automotive parts · Molding & paint

One appearance standard instead of many opinions

Challenge

Pass/fail on scratches, dents and paint blemishes varied by inspector, driving both customer claims and overkill waste.

Approach

Sensitivity maps calibrated to limit samples, with heatmap evidence attached to every verdict, formed a standard inspection scheme.

Result

Criteria unified as data: claim responses got faster, and good parts previously scrapped as overkill were recovered.

Film · Continuous sheet

Full-width, full-length monitoring with line-scan

Challenge

Pinholes, particles and streaks on high-speed sheet were managed by eye and sampling, producing roll-level losses.

Approach

Line-scan full-width imaging combined with anomaly heatmaps, logging each defect position in meters along the web.

Result

Defect positions form a roll map, so downstream removes only affected segments — whole-roll scrap became partial loss.

* Figures and scenarios reflect representative deployments and may vary by process and data conditions.

/03 Specifications

Specs & Integration

Imaging
Area (20MP+) · line-scan (8k/16k); process-specific lighting and optics included
Training data
200–1,000 normal images recommended; no defect labels required
Verdict speed
Under 50ms per image (tunable by resolution and deployment spec)
Detected items
Scratch · dent · pinhole · particle · stain · coating defects · thickness variation, etc.
Outputs
Pass/fail · pixel heatmap · defect coordinates/size/grade · roll map (continuous processes)
Deployment
Edge GPU · on-premise server
Integrations
Rejector/marking trigger · MES quality history · evidence-image archive
/04 FAQ

Frequently Asked Questions

Does it really work without defect samples?
Yes. The verdict model is built from normal images only. Defect samples are used in small numbers just to validate thresholds — and if you have none, we validate with defects collected in early operation.
We worry about overkill (good parts rejected).
It is the key managed metric in early deployment. Region-wise sensitivity maps, size thresholds and limit-sample calibration keep overkill at target levels — typically under 2%.
What happens when the product changes?
A re-collection of normal images and retraining handles the changeover. With the same imaging setup, new-product support typically takes days — a fit for high-mix lines.
How does it relate to our existing AOI?
Both replacement and complement are possible. The most common setup places it behind rule-based AOI to catch unstructured defects; new lines deploy it standalone.
Get Started

Send 100 sample images — we will show you what is possible

Tell us about your process and data environment — we will come back with a feasibility summary and expected impact.