April 2020

51 www.drivesncontrols.com April 2020 MACHINE VISION n Will deep learning enhance or replace vision systems? O ver the past decade, we have witnessed significant technology changes in areas such as big data, AI (artificial intelligence), the Internet of Things, robotics, blockchains, 3D printing, device mobility and machine vision. In each of these areas, novel things came out of r&d labs to improve our daily lives. Engineers are adopting and adapting these technologies to cope with their tough operating environments and constraints. Strategically planning to adopt some or all these technologies will be crucial to the future of the manufacturing sector. Let’s focus here on AI, and specifically deep-learning-based image analysis – also sometimes known as example-based machine vision. Combined with traditional rule-based machine vision, it can help robotic assemblers to identify parts rapidly and precisely, and to detect if parts are missing or have been assembled incorrectly. GPUs (graphics processing units) gather thousands of relatively simple processing cores on a single chip. Their architectures look like neural networks. They can be used to deploy biologically-inspired and multi- layered “deep” neural networks which mimic aspects of the human brain. Using such architectures, deep learning can be used to solve tasks without being explicitly programmed to do so. Unlike classical computer applications which are programmed by humans and are “task- specific”, deep learning uses data (images, text or numbers, for example) and trains itself using neural networks. Starting from primary logic developed during initial training, deep neural networks refine their performance continuously as they receive new data. Deep learning is based on detecting differences: it looks for alterations and irregularities in sets of data. It can react to unpredictable defects. Humans do this naturally, but computer systems based on rigid programming are not much good at this. (But unlike human inspectors on production lines, computers do not get tired of constantly having to perform the same iteration.) Typical applications for deep learning that are already widely used include facial recognition (to unlock computers or to identify people in photos), recommendation engines (on streaming services, or when shopping on e-commerce sites), spam filtering of emails, diagnosing diseases, and detecting credit card fraud. Deep learning can produce accurate outputs based on the trained data. It can predict patterns, detect variances and anomalies, and make critical business decisions. The technology is now migrating into manufacturing for quality inspection and other judgment-based uses. When implemented in the right type of manufacturing application, deep learning, used in conjunction with machine vision, can help to boost profits, especially compared to investments in other emerging technologies that might take years to pay for themselves. Complementing vision? Machine vision systems are based on digital sensors in industrial cameras. They acquire images and feed those images to PCs, where specialised software processes, analyses, and measures various characteristics that are used to make decisions. These systems perform reliably when used with consistent and well- manufactured parts. They operate using step- by-step filtering and rule-based algorithms. Deep learning might sound like the latest buzz-phrase in the world of machine vision. But what exactly does it entail, and will deep learning system replace traditional vision technologies? Ruben Ferraz, field product marketing manager for deep learning at Cognex, explains. Traditional machine vision (left circle) and AI-based deep learning systems (right) each have applications for which they are ideal. There are some applications (shown in in the overlapping area) where they both have roles to play.