What AI visual inspection is, how it differs from rule-based machine vision, which defects it does and does not suit, and the steps from sample review to…
How does AI visual inspection differ from conventional machine vision inspection, and how do we judge whether it fits our line?
Conventional machine vision decides by rules people set (brightness, size, position); AI visual inspection decides with a model trained on images of good and defective products. It suits defects such as scratches and foreign matter, whose varied shapes are hard to define in rules. But if a defect is not visible in the image, AI cannot find it either, so check the imaging first and decide after testing with samples of your own product.
In one sentence
AI visual inspection is inspection in which a trained model looks at product images taken by a camera and separates good from defective or finds where a defect is. It is used where people inspect by eye, or where rule-based machine vision struggles to separate good from defective.
Camera and lighting come first, AI after that. A defect must appear in the image for the model to learn it.
How it differs from rule-based inspection
Item: Basis for decisions; Rule-based machine vision: Values set by people: brightness, area, length, position; AI visual inspection: Criteria the model acquires from training images
Item: Good at; Rule-based machine vision: Decisions that can be set as values, such as dimensions, presence or position; AI visual inspection: Visual defects of varied shape, such as scratches, foreign matter and stains
Item: What to prepare; Rule-based machine vision: Decision rules and thresholds; AI visual inspection: Images of good and defective products, and labels
Item: When a new product or defect appears; Rule-based machine vision: Change the rules; AI visual inspection: Add images and retrain
Item: Common problem; Rule-based machine vision: Slight differences get caught as defects, so false rejects rise; AI visual inspection: Types not in the training can be missed, so a retraining procedure is needed
You do not have to choose only one. It is common to use both on one line, measuring dimensions and checking visual defects with AI. IntLearn Pro also includes a dimensional measurement (IntMeasurement) module alongside its AI engines.
What it suits and what it does not
These are the cases worth reviewing AI for first.
Defects an experienced inspector knows on sight but that are hard to turn into rules
Lines where tightening the rules raises false rejects and loosening them raises missed defects, so thresholds keep getting adjusted
Inspection where the decision criteria waver along with the person on duty
These are the cases worth looking at other methods for first.
When defects are not visible in your current images: look at the lighting and camera again first.
Dimensional inspection that ends with measuring a tolerance: measurement is simpler.
When there is no internal agreement on what counts as a defect: if the criteria are not aligned first, the labels will waver too.
Can we start with few defect samples?
The rarer defects are on a line, the more often this question comes up. There are several ways, and which one is right depends on seeing the product.
Anomaly detection (IntAnomaly): finds parts that differ from the usual product.
Defect synthesis (IntSynthesis) and data augmentation (IntAugmentation): increase the training images.
IntLearn Kinesis: builds robots, conveyors and cameras as a virtual cell and trains on synthetic data, to check imaging conditions and inspection performance before installation.
Even so, real defect samples make test results far more trustworthy. A test without defect samples cannot calculate a missed-defect rate.
Adoption follows this order
Consultation and sample review: we review the inspection target, the defects to find and the line conditions with you.
Pre-install validation: imaging conditions and inspection performance can be checked in an IntLearn Kinesis virtual cell.
Line deployment: images captured with IntCloud Optic train IntLearn Pro, and the inspection package (.izpkg) is applied to IntScan Pro or IntSight.
Review and retraining: results are reviewed, and data from the line is used to retrain and revalidate the models.
If you already have equipment, step 1 starts with its saved images. The details are in what to check when adding AI to an existing inspection machine.
How are results judged?
A single accuracy number is hard to judge by. Look together, on the same samples, at defects that got through (missed defects), good parts caught as defective (false rejects) and whether it keeps up with line speed (throughput). Also decide first whether the evaluation unit is the image or the product. How to check missed defects, false rejects and throughput sets this out, along with a table format.
Limits and cautions
This article is a general explanation. Whether it can be applied to your product can only be known by reviewing samples, imaging conditions and the process. The same goes for whether the lighting and camera can stay as they are, and which engine is right.