Visual inspection for high-mix floors, implemented with Nsight Edge and Vision AI.

Target sites

We have in mind visual-inspection floors that are high-mix, change over frequently, and where defect samples are hard to collect.

• Lines with many SKUs and frequent switching — floors handling many SKUs where switching happens often mid-production.
• Low-volume high-mix floors where defect samples are scarce — small lots, so not enough defect samples for training.
• Floors that rely on manual visual inspection and depend on individuals — appearance judged by the human eye, with results that vary by inspector.
• Floors that want to retrofit inspection onto an existing line — adding visual inspection to a line or equipment already in operation.

Common pain points

• Per-SKU setup can't keep up — each SKU needs its own inspection settings and tuning, so standing up can't keep pace.
• Not enough training data — few defect samples, so the data needed for training doesn't accumulate.
• Changeover is costly — every switch incurs setting changes and training cost.
• Accuracy and speed vary by inspector — manual inspection means accuracy and speed differ from person to person.

Limits of conventional methods

Rule-based alone, conventional deep learning alone, and manual inspection each hit limits on floors with many SKUs and few defect samples.

• Limit of rule-based alone — conditions must be built per SKU, so settings and tuning balloon as SKUs grow.
• Limit of conventional deep learning alone — a large amount of training data is needed per SKU, hard to collect in low-volume high-mix.
• Limit of manual inspection — judgment depends on individuals, training burden is heavy, and accuracy and speed break down in busy periods.

How Nsight solves it

We design from input through integration as a single flow, and implement visual inspection on Nsight Edge. Production good/defect judgment is anchored on rule-based and conventional deep learning, while the VLM is used to support training-data creation.

Products & engines used

Deployment patterns

• Retrofit onto an existing line — add visual inspection to a line or equipment already in operation.
• From optical design — design cameras, lighting and lenses to fit the inspection target.
• Training-data support for low-volume high-mix — NG-image generation and auto-annotation keep training-prep effort low even with few defect samples.
• PLC / line-control integration — connect judgment results to PLC and line control.
• From PoC to full implementation — validate first with a few SKUs, then proceed to full implementation in stages.

Inspection items we handle

• Scratch / dent / chip detection — surface scratches, dents, chips and similar appearance defects.
• Dirt / foreign-object detection — surface dirt and foreign-object adhesion.
• Print / label inspection — presence, fading and errors of print and labels.
• High-mix / low-volume high-mix inspection — visual inspection for floors with many SKUs and small lots.
• Handling SKU additions and changeover — inspection built to match added or switched SKUs.

Built on Nsight Edge and Vision AI

Multi-SKU visual inspection isn't a standalone tool — it is implemented per floor on Nsight's core product, Nsight Edge, and the image-inspection engine that runs on it, Vision AI. By designing not only the AI model but also the optical design of cameras, lighting and lenses, and line integration, it becomes visual inspection that works on the floor. Production good/defect judgment is handled by rule-based and conventional deep learning, while the VLM plays the role of lowering training cost through NG-image generation, auto-annotation and the like.