AI Trend Research

← AI glossary

AI glossary · Basic · Computer vision · Manufacturing

Vision inspection

Vision inspection is a quality inspection method in which AI looks at camera images of products and automatically finds defects such as scratches, foreign matter and assembly faults.

About this term
Level
Basic · Technical terms you often meet in AI news
Category
Industry AI
Industries
Manufacturing
Added
2026-10-10
In plain words It is like having an assistant with tireless eyes scan the products first, instead of an inspector staring at them all day, and pass only the unclear cases to the inspector. However, this assistant can miss new kinds of defects it has never learned about.

How it works

Cameras and lighting are installed on the production line to photograph products at a consistent angle and brightness. After preprocessing to clean up the images, a deep learning vision model judges whether there is a defect and, if so, where it is and what kind it is. Based on the result, good products pass through, and suspected defective products are pulled aside for a worker to check again. Defect images confirmed by workers are collected and used to retrain the model.

Vision inspection lineCamera capture → Preprocessing; Preprocessing → Vision model; Vision model → Good product passes; Vision model → Suspected defect; Suspected defect → Worker recheck; Worker recheck → Defect image collection; Defect image collection → Vision model(Retraining)Camera captureFixed lighting · anglePreprocessingVision modelDefect location · typeGood productpassesSuspected defectWorker recheckDefect imagecollectionRetraining
Vision inspection line Captured images go through the model's judgment and are sorted into good products and suspected defects. Defect images confirmed by workers are used for retraining (dashed line).

Why it matters

Visual inspection by people tends to give inconsistent results because of fatigue and differences between individuals. Older rule-based image inspection required criteria to be set one by one, such as 'this much brightness difference means a defect,' so it struggled with defects of varied shapes. AI vision inspection learns what defects look like from example images, so it can more easily handle complex surfaces and many types of defects. Because it can inspect every unit quickly, it is used to reduce defective products reaching customers.

Things to know

Defective products are much rarer than good ones, so there are often not enough defect images for training. For this reason, anomaly detection, which learns only from normal images and flags anything different, or data augmentation is sometimes used as well. Performance can drop when lighting, camera position or raw materials change, so regular checks and retraining are needed. For safety-critical products such as auto parts or medical devices, inspection records should be kept, and process standards should make clear who is responsible for missed defects.

Adoption and operations checklist

Data

  • Are there enough defect images?
  • Are labeling criteria the same across inspectors?

Operations

  • Is the system revalidated when lighting or cameras change?
  • Do people check suspected defects?
  • Are inspection records kept?
Items to check when bringing vision inspection onto the factory floor and maintaining it.

Example

In an electronic components factory, a camera photographs each circuit board as it passes along the line, and AI marks boards suspected of soldering defects or missing parts. Workers recheck only the marked boards, and photos confirmed as actual defects are added to the model's training data.

Related terms

CNN Anomaly detection Data labeling Data augmentation Class imbalance Smart factory