How Deep Learning Fits into the Automated Inspection Toolbox

Few buzzwords have loomed larger over the machine vision and industrial automation space than “deep learning.” Somewhere around seven years ago or so, the term rode in on a massive wave of marketing hype, with words like “revolutionary” and “disruptive” attached as descriptors. Several years later, the dust has settled and the role of deep learning in automation and manufacturing has become clearer. 

Deep learning, of course, is not magic. It is not going to solve all automated inspection needs, nor is it appropriate technology for all types of applications. It is also not a standalone technology. Rather, it’s a complementary technology to traditional machine vision and automated inspection technologies, and it’s there where deep learning has settled into a quieter but still quite valuable role. 

Real-World Deep Learning Use Cases

Before even considering deep learning for an application, it’s important to understand how it works. Where discrete, rules-based machine vision algorithms are programmed using specific numerical inputs into traditional arithmetic algorithms and convolutions, deep learning software “learns” by analyzing data (images) that have been labeled and categorized by human subject matter experts. Labeling and categorizing defects are crucial parts of the process, as that helps produce “good data” from which the software can create a mathematical model based on small image variations and other clues that a human might otherwise use to decide if a part is good or bad.

While deep learning is not well suited for applications involving precision measurements, location, or anything involving parameterization, it has proven to be a useful tool for making subjective inspection decisions, especially on parts that are highly variable. It is also effective when it comes to inspecting highly complex scenes, where identifying specific features may be difficult. The simplest way to think of it, however, is like this: In any inspection where a human could classify either a defect or a good part, then a deep learning system can likely be applied if it is properly configured.  

But what does that mean in a real-world deployment? Let’s take a simple example of trying to identify flaws or defects in textiles or fabrics such as leather. Trying to program traditional machine vision algorithms to identify these flaws is extremely difficult, but with deep learning, it’s rather straightforward. Deep learning algorithms can learn the varieties of leather and what the textured background looks like over a wide sample set, and make effective decisions based on that training. 

Another common example is weld inspection, an application where welds done on metal will all have varying shapes and sizes but may still be perfectly acceptable parts. Traditional machine vision would struggle immensely with inspecting the different number of welds on an assembly line, but deep learning can be quite effective due to its ability to compare images of defective welds to acceptable welds. Of course, many other applications and industries exist where deep learning may be useful, including semiconductor and electronics inspection, pharmaceutical and medical manufacturing, automotive manufacturing (including electric vehicles), food and beverage, and more. 

Continuous Improvements 

With a proper understanding of deep learning—and when all factors have been considered and all steps have been followed to implement deep learning technology into a new or existing automated inspection system—it can deliver significant value on the manufacturing floor. Common tasks today include feature classification, defect detection, assembly verification, and inspecting scenes where identifying certain features can be difficult due to variability or complexity, such as aiding in vision-guided bin picking applications. 

Moving beyond the hype of deep learning, we see that it has established itself as something different than the magic solution it was initially lauded as, and manufacturers today are reaping the benefits. While it is certainly not right for every application, deep learning does, and will continue to, offer real-world benefits in situations where continuous improvements in inspection accuracy are a top priority.