Machine Vision and Deep Learning in real world application
Let’s consider a couple of projects that I’ve had visibility on that involved deep learning from the start of the project through the end of the implementation. One is an application involving a bottling process, a capping process specifically, and the important thing to think about is the start of the process always is analysis of the application, understanding the needs of the application. With a deep learning workflow, a follow-on to that once the specifications have been set in stone, is to gather suitable images, sufficient images for creating a training data set, the initial data set to train the model that’s going to analyze the product. This is a big step and bigger than a lot of people think. Really, when implemented correctly, it involves installing the final imaging system with the plant floor, allowing that imaging system to capture live production images, allowing it to acquire a sample set of suitable defects, whether they’ve been manually produced or even produced in the production process, and then the workflow shifts to competent labeling of the images. And this is really the process that we undertake as an integrator for example, to do this all the time. Collaborative labeling of the images to identify the defects or the variations that would cause the part to fail, and testing that data set and making sure it’s reliable. In the application that I have in mind, that involved imaging a bottle cap under thermal imaging to detect the release of the inner, or the failure of the inner foil to be sealed against the surface of the cap by an induction heating system. Variations in the profile of that heat map coming from the thermal camera indicated that the seal was not safe, was not solid, and caused the bottle to be rejected. And after initial tuning, and then subsequent, which is another part of the workflow, subsequent acquisition of images and re-tuning the quality of the data set, that system provided excellent results that actually caused the manufacturer to change some of their processes and tune some of their production to yield better quality product to the customer.