machine vision eye

How Deep Learning Enhances Machine Vision in Manufacturing

What Is Deep Learning?

Deep learning is a subset of machine learning, which is itself a branch of artificial intelligence (AI). Artificial intelligence is, as John McCarthy, one of the founders of AI said, “the science and engineering of making intelligent machines“. Therefore, it stands to reason that deep learning provides the advanced analytics tools to handle the massive amounts of data that result from the deployment of sensors and the Internet of Things (IoT) in “smart” manufacturing. As it applies to machine vision software, deep learning allows machines to learn from enormous sets of images.

To aid a machine in carrying out a particular task, a deep learning program utilizes many layers of software neural networks–which are designed like the neural networks in the human brain–to learn the requirements needed to perform the task. Humans train deep learning programs to acquire information much in the same way children learn. Over time, the programs learn to distinguish “good” from “bad” patterns after viewing thousands of images designated as such through tagged product examples.

How Does Deep Learning Differ from Machine Vision?

Machine vision software inspects consistent, well-made components dependably because it employs rule-based algorithms and step-by-step filtering based on their characteristics. It tolerates some variability in the appearance of a component because of scale, rotation, and distortion of the component in its position and orientation in space.

But image quality problems and complex surface textures create serious challenges in the inspection process. Machine vision systems often have difficulty sorting out variability and deviation between parts that very closely resemble each other. While “functional” deviations are usually caught and rejected by machine vision systems, cosmetic deviations may not be since these systems may not be able to distinguish them.

A machine vision system is also severely limited in final assembly inspection because of variables that can be difficult to single out such as lighting, color changes, curvature, and field of view. For this kind of inspection, deep learning is much more effective.

In contrast to machine vision, deep learning programs work best for inspection applications with no predefined shape. Finding cosmetic imperfections such as scratches or locating an object with a predefined shape are tasks that deep learning programs perform very well. These programs can also simplify packing and validation by confirming that all components are correct and in place. The downside of deep learning is that processing the information about an object can sometimes be rather slow because of all of the data that has been gathered and learned, so machine vision’s higher processing speed gives its a decided advantage.

What Are the Implications of Deep Learning for the Future of Machine Vision?

Deep learning programs excel in the inspection of irregular or anomalous objects and have the granularity to sort out variables in final assembly inspections. These capabilities expand the reach of what a machine vision system can do, making these two technologies complementary. Although these systems require a “human in the loop” to train and to tag product examples, machine vision and deep learning together far exceed the capabilities of human inspection in consistency, reliability, and speed. Deep learning has the potential to scale and because that potential has only begun to be realized, it promises to expand the scope of capabilities in the manufacturing industry.

For more information about machine vision solutions that incorporate deep learning techniques, contact Integro Technologies today.