Is There an Advantage to Moving Machine Vision Systems to the Cloud?
That depends on how you want to use your system and who you want to be able to operate and maintain it.
“How do you see AI and cloud technologies having an impact on computer vision and machine vision?”
I was recently asked this question on an episode of the Industrial Automation Insider podcast.
I immediately called out the need to look closer at cloud-based training of the AI models that serve as the decision engines within these automation systems.
Why haven’t more manufacturers, warehouse operators, or healthcare providers jumped on this opportunity? They’re all using machine vision systems in various ways, often across multiple sites that are geographically distant from one another.
As discussed in a Forbes Tech Council article, when you move the training of deep learning neural networks or other AI to the cloud:
· You don’t need someone with a PhD in optics on site, or at every site, to install a camera.
· You can democratize machine vision in a way that benefits the entire workforce and the people you serve through your business.
Why the delay?
Well, as the podcast host later called out (off the air), there will likely be resistance or hesitancy to move vision systems – or at least the training of such systems – to the cloud for a few reasons:
· Overcoming the fear of change is as difficult as overcoming the fear of failure.
· There are fears of job loss if the project results in cost overruns or fails to deliver anticipated cost savings – to include potential labor cost savings.
· Sometimes what you’ve got in place already is good enough.
That begs the question: How do you know if you should move to a vision system that would allow for cloud-based training of deep learning AI models? Is there a certain set of conditions in which it’s lower risk to move to the cloud than it is to stay on-premises?
Start by answering these questions:
· Do you have several manufacturing lines in different factories around the world that produce similar goods? Or perhaps several locations at which surgical tools are being sanitized and packaged for reuse?
· Are your computer vision or machine vision applications challenging to the point where on-premises training models would be prohibitive? Would it be impossible to use a rules-based traditional programming model? (For example, searching for scratches on an object, which have zero uniformity.)
· Do you want to simplify the deployment of your existing machine vision or computer vision systems?
If you said yes to any of the above, strong consideration should be given to cloud-based AI training.
A Word of Caution as You Move to the Cloud
I’ve seen many manufacturers underestimate the complexity of deep learning AI models, especially those used to train machine vision systems. I’ve also seen them miscalculate the cost, but only because the high “cost” of using deep learning AI comes from not fully understanding how it works.
The scariest thing that I hear is this:
“Oh, I'm so interested in deep learning. I'm going to go get an intern for the summer and annotate all my 275 SKUs.”
That is not a good way to start.
You need the expertise of quality control people who have seen these parts for years and years, as they know what is good and what is not good.
You also need data scientists who can help you really understand how you need to model this and what kind of percentage of good parts versus bad parts you want to have in your training set.
Additionally, you need a data scientist who generates synthetical images based on the few defects that you have.
If you have that understanding, that knowledge, then your ability to create models and to scale is very high.
If you don't, then you'll fail, and everything will be very, very difficult.
So, it’s very important to understand the value of entry.
Many people think that this can be done very quickly and very easily because it’s AI. They think it doesn’t cost any money. That’s part of the reason why 80% of people struggle to deploy deep learning AI in their machine vision systems. Those things simply aren’t true.
The other reason why people struggle with deep learning AI: they try to start with a very complex problem. Many say, “Let's revolutionize our factory and be 100% AI based.” Yet, all they need to do is find scratches or holes in products coming down the line.
So, if you decide to use deep learning AI in your machine vision or computer vision system and opt to train it via the cloud:
Build the system one use case at a time.
Find one problem that you want to solve that is very specific – that is easy to describe and contain.
Make sure you can measure the output of it. (If you’re using the deep learning AI machine vision system for quality control, you can usually measure pretty well.)
Then build a team. That doesn’t mean you must hire a team. Simply work with partners and suppliers who can come in and help you out so that you start gaining confidence in this.
