Drives and Controls Magazine
Deep-learning machine vision opens up new possibilities
Published:  04 February, 2019

The emergence of machine vision technologies that use “deep learning” is expanding manufacturers’ capabilities and flexibility, leading to greater cost efficiencies and higher production yields, according to a new report from ABI Research. It predicts that these technologies will achieve a CAGR of 20% between 2017 and 2023, with revenues reaching $34bn by 2023.

While conventional machine vision systems are easy to implement, they are limited in their capabilities. They rely on pre-programmed rules and criteria that support a limited range of functions. Deep-learning-based machine vision, by contrast, can be trained and improved.

“This is, in part, driven by the democratisation of deep-learning capabilities,” explains ABI principal analyst, Lian Jye Su. “The emergence of various open-source artificial intelligence (AI) frameworks – such as TensorFlow, Caffe2 and MXNet – lowers the barrier to entry for the adoption of deep-learning-based machine vision.

“In the past, the choice of machine vision solutions was limited to a handful of companies that performed relatively simple image-processing operations,” Su adds. “With deep-learning-based machine vision, manufacturers can opt to develop their own deep-learning-based machine vision systems without the worry of vendor lock-in.”

As well as cameras, deep-learning-based vision systems can also incorporate data from other sources, such a lidar, radar, ultrasound, and magnetic field sensors, providing insights into other aspects of production processes. Compared to conventional machine vision which can only detect product defects and quality issues defined by humans, deep-learning algorithms can go further. They can pick up unexpected product abnormalities or defects, providing flexibility and insights to manufacturers.

The machine vision giant Cognex entered the deep-learning market when it acquired the Swiss developer of deep-learning software, ViDi Systems, in 2017

Deep-learning-based machine vision needs a robust cloud platform that supports condition-based monitoring, sensor data collection and analytics. Unlike conventional machine vision which relies on line-by-line coding, deep-learning-based machine vision models can be deployed without significant coding experience, because they achieve unsupervised learning based on data gathered.

“Manufacturers are still opening up to adopting AI capabilities into their workflow,” Su concludes. “Deep-learning-based machine vision will serve as the right catalyst to move the needle, as the potential is enormous. Start-ups that start off as deep-learning-based machine vision solution providers, are also starting to enable big data processing, process optimisation and yield analytics.”