Can you describe the defect in numbers?
If the defect is a gap over a size, a missing piece or a wrong code, a rule usually catches it with very few false rejects.
Guides
Rules or deep learning for this defect?
Deep learning is the right tool for some defects and the wrong one for others. The defect decides, not the buzzword.
What is the difference in practice?
Rule-based vision measures things you can describe: an edge position, a gap, a color, a character. Deep learning learns from labeled example images and judges new images by similarity. Rules are easy to explain and adjust. Models handle variation that no rule can describe, but they need data and care to keep trusted.
How do you decide?
Does good product vary a lot?
Natural products, wrinkled film and shifting backgrounds make good packs look different from each other. That is where models earn their place.
Is the lighting stable?
Stable light favors rules. When light changes and cannot be controlled, a model trained on that range copes better.
Do you have labeled images?
A model needs good and bad images, labeled by people who agree. If bad parts are rare, collecting them can take months.
Who keeps it running?
Rules can be adjusted by a trained technician. A model needs a process for adding images, retraining and re-testing after changes.
Can you combine them?
Often the best station uses rules for position and codes and a model only for the one defect class that varies.
Start with the simplest check that catches the defect on your samples. Add a model only for what the rules cannot see.
FAQ
Common questions
QIs deep learning more accurate?
Not by default. On defects a rule can describe, a rule is often as accurate and more predictable.
QHow many images does a model need?
It varies by defect. Some models start from a few dozen good images, others need hundreds of labeled defects.
QIs a model harder to change control?
It takes more steps. Each retrain should be tested against a fixed set of samples before it goes back on the line.
QCan we start with rules and add a model later?
Yes. Many stations do exactly that once enough defect images have been saved from production.