When the concept of “thinking machines” emerged in the 1950s, alarmist warnings about this new field of artificial intelligence soon followed. The fear of the rise of the machines has played out in pop culture ever since, from the iconic 1968 film “2001: A Space Odyssey” to the more recent “Ex Machina.”
While AI hasn’t yet taken over society, improvements in data storage and processing power have enabled cognitive systems like IBM Watson that are designed to take the guesswork out of human decision-making. Most current iterations of AI, however, tackle more modest tasks like object recognition.
The promise of AI looks to enable machine vision to take on challenging applications beyond the capabilities of today’s solutions. But is the technology ready for primetime in industrial applications?
Testing the AI Waters
AI’s applicability in machine vision relies on the affiliated branches of machine learning and, more so, deep learning. At its broadest level, AI is a computer’s ability to simulate human intelligence. Diving deeper, machine learning gives computers the ability to act without being explicitly programmed. Deep learning, a subfield of machine learning, enables computers to learn from experience.
Several developments over the last decade have made deep learning a reality, not just a possibility, for machine vision. “Based on new techniques in neural networks, sufficient computational power in GPUs, and an abundance of data, only now can we use artificial intelligence for image processing,” says Olivier Despont, Business Development at ViDi Systems, a Swiss company that makes deep learning–based vision software.
Deep learning holds promise over traditional machine vision because, unlike traditional image processing software that uses a rules-based approach, “AI is the next step where we take things that are not easily characterized or non-linear and give them to the machines to create that next level of repeatability,” says Wallace Latimer, Sales Director, Customized Optical Systems at FISBA LLC (Tucson, Arizona).
“Whereas linear algorithms create a very narrow bucket, AI/deep learning creates bigger buckets that can accept more variation,” Latimer continues. “It’s widening the acceptance band of what is good or bad, and why it’s good or bad. By having the bigger bucket, you can focus on what offers the biggest bang and reduce changes to inputs.”
At least one deep learning system is on the market for machine vision users today. ViDi Suite from ViDi Systems is the first commercially available deep learning–based industrial image analysis software. The software, which integrates with standard image processing libraries, learns like a child does.
“You don’t teach a child using a rules-based approach by explaining what a house is,” Despont says. “Based on few examples, our brains, even at a young age, are able to extract what makes a house. Our system works exactly the same as the human brain.”
ViDi Suite, which has won nearly 20 awards, comprises three different tools. ViDi Blue finds and detects single or multiple features within an image. The tool localizes and identifies complex features and objects by learning from annotated images. ViDi Red detects anomalies by learning the normal appearance of an object, including its variations. The red tool also segments specific regions in images. ViDi Green learns to separate different classes based on a collection of labeled images to classify an object.
In addition to new machine vision solutions that can handle greater product variation, another advantage of deep learning over traditional machine vision solutions is that it can reduce the time necessary to develop a machine vision program. “With the classic vision approach, many applications need 60-plus days of software development and feasibility,” Despont says. “ViDi can complete development in half a day.”
Unlike Facebook, Google, and IBM AI systems that use server farms to power their software, ViDi uses a single high-end NVIDIA GPU to train the system in a matter of minutes, rather than the days or months it takes to program and parameterize with IBM Watson, according to Despont.
“And instead of using millions or billions of images, we recommend starting with 30 to 50 representative good images to teach the system,” Despont says. “We’re not sending images to a cloud-based server farm to do the processing or training. Customers are happy they can run everything on a single PC with one GPU and keep ownership of their images.”
Opportunities and Challenges
Deep learning shows particular promise in applications that present challenges to traditional vision systems. “AI is really suitable in food inspection among others where you want to inspect donuts or a piece of meat that shows significant difference from one instance to another,” says Bruno Ménard, Software Program Manager at Teledyne DALSA (Montreal, Québec).
But it’s not just organic inspection applications that will benefit. Ménard cites traditional defect detection applications as another example. “It’s difficult to program a computer with traditional algorithms to define the defect without having to redo the settings every time there is a new defect,” he says. “But by using artificial intelligence with a lot of samples, you can end up with a really good definition of what is a good part and what isn’t.”
As AI emerges in machine vision, the technology will find a place in additional inspection tasks and eventually extend beyond the realm of industrial automation. According to FISBA’s Latimer, deep learning will be advantageous in markets such as medical, life sciences, food, counterfeit inspection, and lumber grading.
“These are areas that all have very gray decision points,” says Latimer. “Is that apple good enough or not? That’s hard to make a linear rule to say it is. Deep learning should enable a lot of applications to become much more efficient and repeatable.”
For his part, ViDi Systems’ Despont foresees that deep learning will include medical diagnostics, surveillance, autonomous vehicles, and smart agriculture for inspection or map analysis. “AI is the future and will be helping people solve some complex tasks very quickly as computational capabilities are doubling almost every one-and-a half years,” Despont says.
Many machine vision professionals recognize the promise that AI and deep learning offer to the vision industry, but they say AI’s full potential won’t be realized for at least another 3 to 5 years. What’s more, AI isn’t necessarily the solution for everything that ails traditional vision and image processing.
Teledyne DALSA’s Ménard notes two major drawbacks in AI systems. “First, you need a lot of training for it to happen, and you need to create the expert to reach the next level of classification,” he says. “The second drawback is once it’s trained and the classification fails, it’s difficult to fix the problem. You have no choice of retraining with a new sample.”
Before artificial intelligence becomes commonplace in machine vision, industry experts believe the industry will have to let much bigger players do the heavy lifting. “From our niche segment, we’re getting to watch the Googles of the world drive this technology to incredible levels of investment and refinement,” says FISBA’s Latimer. “Our industry cannot invest the time and money at the necessary scale. We’re going to have to leverage it.”