OKMeasurements and rules
When the defect can be described with rules: measurements, presence, character reading. Fast, stable and easy to explain.
Home / Machine vision / Deep learning
Machine vision · artificial intelligence
Deep learning in inspection: where classic vision falls short
Some defects cannot be described with rules: a seal with folds, a surface with reflections, a natural product that is never the same twice. For those we use deep learning, with neural networks trained on your own products and integrated into the line.
Our approach
Classic vision, deep learning or both
Not everything needs artificial intelligence. In each case we work out which technique gives the reliability you need and tell you before you sign, including anything that can't be guaranteed.
Difficult defects
For whatever varies from one pack to the next: seals, folds, reflections, surfaces, materials or natural and highly variable products.
The process in numbers
Histograms and trends for each measurement, to see whether the process is drifting out of adjustment, before the first reject.
A real case
Bottom seal, monitored by deep learning
With our Nadir software we detect sealing problems on the bottom of carton packs using deep learning, and process statistics anticipate drift before it turns into defects.
- Neural networks trained on your good and bad packs
- Mean, spread and bell curve for each measurement
- Drift alarm, before the first reject
- Photo of every reject and instant alerts to whoever you choose
How we work
From sample to line
Deep learning is part of our applied artificial intelligence: concrete tools, measured by what they improve on the line.
With your products, at no cost
You send us good and bad parts and we tell you whether the defect can be detected in line, with which technique and with what reliability to expect.
A network for your product
A few dozen good and bad parts are enough to start training. If there isn't enough data, we tell you.
Integrated and measured
The model works on the line as one more tool in the recipe, with the photo and the reason for every reject.
FAQ
What we're often asked
When is deep learning needed in an inspection?
When the defect cannot be described with rules: seals, surfaces, folds, reflections or natural and highly variable products. If the defect can be measured, classic vision is simpler and more stable.
How many images do I need to start?
Not always many: a few dozen good and bad parts are enough to start training. We tell you honestly whether there is enough data or what is missing.
Do my images leave the company?
That is your decision. We can work with models that run on your premises or with cloud services under contract, and we explain it before we start.
Can it be combined with classic vision?
Yes, and that is the usual approach: rules for what can be measured and deep learning for what varies, in the same inspection.
Free feasibility test
Does your defect refuse to be described with rules?
Send us good and bad samples. We'll tell you which technique works with your product and with what expected reliability.