How do vision systems make decisions based on what they see?
While rule-based machine vision is still the prevalent technology, various types of AI-powered machine learning have become capable and flexible enough to take over in many applications.
Often, a combination of rule-based and AI-powered machine learning can provide the most efficient solution.
Rule-based systems
Traditional rule-based machine vision uses specific if-then rules to make decisions about an image. “If the bottle is filled to at least this level, let it pass.” The rules are programmed by vision engineers with deep knowledge of the best combination of techniques to get the desired output.
The rules are used to create programs that tackle specific jobs, such as detecting the edge between light and dark areas or measuring the distance between two points.
These rules can then be used one after another on an image, in what is called “tool chaining,” to perform a wide range of sophisticated tasks. They’re also very good at high-speed, high-accuracy inspections on products or parts that are predictable and consistent.
Deep learning
Deep learning uses AI to power machine vision systems by using examples to train the software with labeled images until it can make distinctions on its own.
This training mimics how humans learn, helping the system learn how to make accurate decisions without being thrown off by irrelevant variation.
Operators train deep learning systems with hundreds of labeled images, such as a variety of possible defects in a manufactured part, or several different assemblies that need to be classified.
One significant improvement over rule-based systems is that deep learning systems learn to distinguish between a wide range of real defects and merely cosmetic variations without needing to be shown every possible result – something time-consuming or even impossible to accomplish using rule-based programming.
In contrast to rule-based systems, deep learning systems don’t require experienced machine vision programmers or expertise to set up. They do need someone with good knowledge of the products being inspected to provide hundreds of images of both defective and good components.
Deep learning excels at assembly verification, defect detection, and classification of complex parts that vary unpredictably. It’s also good at providing accurate results even when parts are reflective or images are distorted.
Edge learning
Edge learning is another type of AI, optimized to meet the needs of industrial automation. It comes pre-trained to solve the types of problems typically faced by industrial automation.
As a result of the pre-training, an edge learning system can be trained on a specific industrial inspection problem with as few as five images, making it even easier and faster to deploy than a deep learning system.
Edge learning enables line engineers to quickly implement an optimized classification or defect detection function on their line, without specialized training. Since it doesn’t require a sophisticated processor, it can be quickly deployed where it is needed.
The advantages of edge learning lie in its speed, little reliance on computing resources, and flexibility, as well as its user-friendliness.
Comparing rule-based and AI-powered systems
The decision between rule-based, deep learning, and edge learning systems mainly comes down to the type of application you’re implementing.
- Rule-based vision tools are effective at a variety of specialized tasks with very consistent targets, such as location, measurement, and orientation.
- Edge learning excels at repeatable tasks with moderately consistent targets.
- Deep learning shines when looking for complex defects or analyzing images with significant variation.