There's a lot of talk about artificial intelligence, machine learning and, above all, deep learning. How and why is machine vision based on these different technologies arriving in industry? What do they offer, and what are their advantages over conventional solutions?
In the manufacturing industry, so-called conventional machine vision solutions have been around for several decades, wherever quality control is necessary and very much in evidence. They are used for presence/absence checks, measurement or positioning when the objects are known in advance and their differences are perfectly identified.
These vision solutions are based on distinguishing rules for a defined defect, which means that these rules have to be multiplied as defects multiply. This can become unmanageable when the product being manufactured presents a very large number of possible variants, without this being considered a quality defect.
The contribution of Deep Learning to machine vision
The advantage of Deep Learning is that we no longer need to provide these specific distinction rules, but will provide meaningful examples.
Using examples, i.e. images taken on the production line, we carry out neural network training, which will then enable us to analyze defects or products we've never seen before.
The approach is therefore to tell the algorithm that such-and-such an image corresponds to such-and-such a defect or perfect product, so that it can then figure out for itself how to process this task to deliver correct results.
Neural network training is therefore carried out upstream of product use, in the Cloud or in the camera itself, to enable subsequent classification or anomaly detection, for example, which are two of the principles of Deep Learning.
Deep Learning to replace conventional machine vision methods
Conventional machine vision is already used for a large number of applications. Many existing technological building blocks, such as shape search or matching for localization, can be used upstream. Deep learning tools complement conventional machine vision tools when these alone become insufficient.
The combination of these different methods makes it possible to solve applications that were virtually impossible to do before.
For example, controlling products or parts made of "living" materials, such as wood, or controlling products in the food sector, were all challenges that we can meet with relative ease today!
In addition to solving complex applications, the configuration effort and skill level required are very low indeed, and accessible to all. To detect anomalies on a product, for example, mechanical installation of a camera, image adjustment (often automatic) and sorting of the images received (very few <20 images) are the only three steps required to implement applications.
The fields of application and industries are very varied, and where conventional methods don't work, the addition of artificial intelligence gets results.
To find out more and discover how simple it is to implement a quality control solution using Deep Learning, watch the Click here to view the webinar produced by SICK.