David Dechow:
I believe, and based on experience and observation, the biggest mistake that end users, systems integrators, users of the technology make when addressing deep learning or AI applications is to ignore the fact that in industrial automation, the task still involves an imaging component. We talk a lot about data being the king in deep learning. It’s the same way in traditional machine vision as well. We need the image data to be quality and to highlight the features that have to be analyzed. So one of the first mistakes that people make in addressing a deep learning application is to mistakenly think that it is going to eliminate the need for quality data. And really that’s not true. So we need quality data on the imaging side, good imaging design, and quality data on the labeling side as well. We need collaboratively labeled data that is done by experts that know exactly what the defect classes are going to look like, and provide that data to the model to be trained. And in combination, I think those are the areas that people fall short on but can very easily be improved.
