Cylindrical-label integrity checking at an OEM
Label text, barcode, and serial — all on a curved housing.
A contract OEM producing cylindrical consumer housings needed to guarantee that every label rolled onto the product had correct text, a readable barcode, and a valid serial number. Classical OCR fails on curved surfaces with inconsistent lighting.
Cylindrical product with label and barcode inspection overlay showing text regions and barcode decode
- Industry
- OEM contract manufacturing
- Product
- Visual Inspection AI Agent
- Scope
- Label integrity on cylindrical housings
- Focus
- Text + barcode + serial verification
A known blind spot for classical OCR.
Label verification on a cylindrical housing is one of the hardest combinations for traditional machine vision: curved surface, reflective finish, variable lighting, and three different kinds of content to verify at once.
Curved-surface distortion
Text on a cylindrical housing is non-planar by definition. Templates calibrated for a flat label lose accuracy as the product rotates through the inspection area.
Reflective and variable lighting
Glossy housings reflect ambient light unpredictably. Classical OCR pipelines are forced to run defensively, trading coverage for stability.
Three verification jobs at once
The line must verify free-form label text, decode the 1D barcode, and validate the serial number format — all in the same frame and within the takt.
Silent escapes are expensive
A wrong serial number on a shipped product is a traceability problem that surfaces downstream, weeks or months later. The customer needed a zero-escape target.
Semantic understanding of the label, not the pixels.
The Insightek team deployed Visual Inspection AI Agent at the labeling station. The agent reasons about the semantic content of the label — text, barcode, and serial — rather than matching pixel templates.
- 01
Register the label semantically
An OK sample is photographed from the inspection viewpoint. The agent identifies the text region, the barcode region, and the serial region — and builds a semantic baseline instead of a pixel template.
- 02
Cross-check content at inference
On every product, the agent runs OCR inside each semantic region, decodes the barcode, and validates the serial against the expected format. A single consolidated OK/NG is reported back to the line.
- 03
Log every annotated frame
Every inspected product is saved with its annotated process image. Trace-back to any unit is a structured log query, not a manual review.
Cylindrical product with semantic label verification overlay
Semantic region understanding is the unlock — the agent is not matching pixels, it is understanding what each label region is supposed to contain.
What the line gained.
This case is covered by an NDA that restricts publication of specific reject-rate and throughput numbers. Below are qualitative outcomes verified with the customer.
Specific reject-rate and throughput numbers from this deployment are NDA-protected. The bullets above are qualitative outcomes the customer has approved for publication. Full figures are available under mutual NDA.
Qualitative outcomes
Label verification coverage closed
Text, barcode, and serial are verified in the same frame, inside the same inference call — eliminating the classical gap where one of the three was out of coverage.
Robust across surface reflectance
Because the agent reasons about the label semantically, reflective housings and lighting drift no longer cause the cascading false-rejects seen in the prior OCR pipeline.
Full trace-back by serial
Every inspected unit is logged with its annotated process image, so downstream trace-back on any serial number becomes a structured query instead of a manual review.
Verify text, barcode, and serial in a single pass.
Book a 30-minute demo. We will walk through your label design, your surface finish, and a realistic POC plan.