在边缘学习和深度学习之间进行选择

Deep learning simulates the way interconnected neurons in the human brain strengthen and weaken connections to create an understanding of images. In deep learning, hundreds of layers of neural networks are exposed to a large set of images of similar objects. By slightly modifying connections within and between these layers every time it is exposed to a new image, deep learning learns to reliably identify those objects, and detect defects in them, without any explicit training.

Traditional deep learning provides the capacity to process large and highly detailed image sets, making it ideal for complex or highly customized applications. Because such applications introduce significant variation, they demand advanced computational power and robust training capabilities. To account for this variation and capture all potential outcomes, image sets numbering in the hundreds or thousands of images must be used for training. Traditional deep learning enables users to analyze such image sets quickly and efficiently, delivering an effective solution for automating sophisticated tasks. While full-fledged deep learning products and open-source frameworks are well designed to address complex applications, the majority of factory automation applications entail far less complexity, making them better suited for edge learning.

The power of AI can be applied to problems in factory automation by embedding knowledge of application requirements into the neural network connections from the start. This pre-training removes a lot of the computational load, particularly when supported by the appropriate traditional machine vision tools. The result is edge learning, a light and fast set of vision tools.

Edge learning tools can be trained in minutes, using as few as five to ten images. Compare this to deep learning solutions, which can require hours to days of training, using hundreds to thousands of images. By streamlining deployment, edge learning enables manufacturers to ramp quickly, while remaining nimble and able to adjust easily to changes.

In order to optimize the edge learning to run on embedded vision systems, the training images are downscaled or fixtured in a way that only the specific region of interest is analyzed. If these downscaled images were to be differentiated with the line engineer’s own eyes, they can be confident the edge learning tools will perform equally as well. However, it is important to note that this optimization comes at a trade-off. It limits the use of edge learning in very advanced and high-accuracy defect detection applications, which are better solved with traditional deep learning solutions.