Machine vision systems are deployed in manufacturing to inspect a wide variety of components, guide robots and perform measurement tasks. However, can they reliably do these checks when the lighting is not consistent, or the parts are not fixtured? Artificial Intelligence (AI) / Deep learning in machine vision can overcome some of these challenges. While AI can be designed to think like humans, unfortunately it is not capable of doing any kind of dimensional measurements.
"Within this industry, I see a major advantage of AI over traditional vision where there is poor contrast in the features or defects and the background"
What the vision industry needs is a blend between AI and traditional vision to solve complex vision solutions.
There’s a lot of development activity going on in the areas of AI and deep learning in Machine Vision. Part features that look different than normal but cannot be defined using pattern matching or any kind of histogram-based tools are a perfect fit for AI. It may not be advantageous to use AI tools to differentiate nuts from bolts as standard pattern matching tools would easily do that. Also, the cost associated with AI runtime licenses is multiples of standard vision libraries. Therefore, it is advisable to use AI on applications that require detection of features or defects that are not easily definable, have poor contrast and do not have well-defined edges or contours.
The exact number of images required to train a model is purely based on the complexity of the part. Some vision software companies believe a typical application might require only 50 or so images. However complex assembly component inspection might require thousands of images. Lighting, part presentation and camera-resolution all play a role in defining image quality which in turn dictates variations that can be compensated for in any given manufacturing environment.
The most challenging part in training a deep learning model is finding sample parts with representative features or defects. For example, if one is designing a new model to detect surface flaws, both good and bad areas should be trained. Developing a deep learning system without enough representative samples is not advisable. As parts become available during initial production runs, they could be added to the model by retraining. Unfortunately, activities that interrupt manufacturing are highly discouraged, therefore such systems are not easy to deploy.
Within this industry, I see a major advantage of AI over traditional vision where there is poor contrast in the features or defects and the background. When inspecting shiny parts like pistons and piston rings, typically diffused lighting technique is used to minimize glare or reflections. Subtle surface defects show very poor contrast in such images and rule-based algorithms do not fare well. However, a deep learning trained model can suit very well in these situations as it can learn based on variations in the surface quality.
In automotive applications for example, an assembled piston goes through more than 25 checks to ensure the presence of components, quality/orientation of components and correct markings on the part. Some of the checks involving dimensional measurements can be easily implemented using traditional vision tools. However, checks like breakage in components, machined surface flaws or missing coating would fare well with deep learning tools. Inspection systems for such parts should be more robust and reliable if both AI and traditional vision are used.
In summary, AI has definitely helped machine vision see more of what it could not easily see before. There are challenges involved in training and deploying AI solutions; however, they are expected to be more consistent and reliable than traditional vision solutions in applications that lack proper definition of features for inspection. Since both AI and vision libraries will at some point be used together to solve complex vision solutions, users will have to determine the best fit based on their inspection tasks and budgets.