Jan Jongboom from Edge Impulse: I would like to see an API that makes it easy to get captured images into the system, regardless of the hardware used to acquire them.

© NurembergFair

"To implement a standard in a system, you usually need additional layers that reduce the performance of the system," explains Jan-Erik Schmitt. "In the PC world, this is less of a problem, as increased performance requirements can still be met by selecting more powerful components, as resource optimization is not usually the focus - in contrast to highly optimized embedded vision solutions."

Another reason for the difficulty in defining suitable standards for embedded vision in all segments, according to the experts, is that embedded vision systems must naturally be more scalable than conventional image processing systems.

Dr. Olaf Munkelt mentions one possible approach that would render many standards obsolete: "Apple recently introduced the M1 processor, a chip that provides extremely high computing power with very low power consumption. This chip contains powerful SoCs for various tasks, which can also mean an enormous increase in performance for embedded vision systems. If all players in the industry were to agree on this direction, it would simplify many problems, especially as Apple also provides the necessary software tools." However, it is unlikely that all embedded vision players will commit to this Apple platform.

Jan-Erik Schmitt from Vision Components: "Today, young people leave university and are already BV experts because they have already dealt with this topic during their training.

© NurembergFair

AI fuels embedded vision

Artificial intelligence methods are currently being used more and more in almost every technical field. The panel also sees this as a clear trend in the field of embedded vision, as Jan Jongboom explains: "From the user's point of view, AI makes it easier to develop systems that generalize well. The use of transfer learning, in which a large number of images already learned from other applications serve as the basis for training a system and only a relatively small number of new images of a specific use case are added, minimizes the effort for users enormously. For example, you can very quickly take a few images of good and faulty parts in production and train the system with them as a supplement. It then independently understands the difference between OK and NOK. This revolutionizes the way in which image processing systems are programmed and dramatically lowers the hurdle for the development of practical models: just a few years ago, it was necessary to collect and train an extremely large number of images."

However, there is also a negative aspect of AI methods, as Dr. Olaf Munkelt points out: "Industrial users usually want an explanation as to why a part is classified as faulty or why a certain decision was made. AI systems are very powerful, but they are not very good at explaining why they made a decision. This statement is completely independent of the topic of embedded vision and means that trust is a core problem when using AI. We invest quite a lot of time in building user trust in these types of algorithms, because users won't use them unless they have that trust."

According to Munckelt, this applies in particular to applications in which very precise results are required: "If there is a larger tolerance range in which inspected parts are rated as good, this makes it easier to gain acceptance for the use of AI algorithms. In other applications, however, this does not work so easily." There is also the fact that AI-based systems can only deliver good results if the previous training was correspondingly good.

Dr. Frederik Schönebeck from Framos: We have to provide architectures that are scalable and still allow easy access.

© NurembergFair

Nevertheless, according to Munkelt, the positive possibilities of AI outweigh the disadvantages, as AI accelerators, among other things, help to simplify the use of embedded vision. MVTec has provided an abstraction layer in its software that makes it easier for developers to work with AI accelerators such as TensorFlow, OpenVINO or other products. "Users really appreciate this because they no longer have to worry about coding all the bits and bytes," says the MVTec CEO.

Is embedded vision too complicated?

Embedded vision generally has a reputation for being a relatively complicated technology. "Plug & play" seemed unattainable in this area for a long time, and in fact the technology was often only suitable for experts due to a lack of tools. Things have changed in the meantime. Jan-Erik Schmitt sees a similar development to that of deep learning: "There are now also tools that are much easier to understand, and a lot has changed in recent years. The tools for implementing embedded vision systems are also evolving because hardware and software are constantly becoming more powerful. This constantly gives rise to new ideas as to where the technology can be used. With the numerous embedded vision applications that have emerged over the past 20 years, there are also more and more people who are interested in this topic and want to delve further into the technology."

Dr. Olaf Munkelt from MVTec Software: Trust is a core problem when using AI.

© NurembergFair

In addition, until a few years ago, mathematicians, physicists or engineers were required to create image processing applications and could only gain the necessary experience through practical use of the technology. "Today, young people leave university and are already machine vision experts because they have already dealt with this topic during their education," says Schmitt. "There are therefore many more image processing experts today than in the past."

In combination with the simplifications that have now been achieved in the use of embedded vision, Schmitt describes the current state of the technology as "plug & smile" rather than "plug & pray". According to Gion-Pitschen Gross, the use of open source software also simplifies the use of embedded vision for users: "Much of it is very easily available today and can be adapted to the respective application with little effort. This trend is still relatively new, but manufacturers such as Allied Vision are now making their drivers available as open source so that users can adapt them to their application and use prepared examples. This also makes it easier for users to implement their applications."

The author: Peter Stiefenhöfer is the owner of PS Marcom Services in Olching.

© PS Marcom Services

Experts agree that the use of embedded vision in industry is benefiting greatly from technological developments in the consumer sector and the increasing computing power of embedded systems. For these reasons, devices equipped with embedded vision, such as mobile barcode or data code scanners, drones, self-driving vehicles, delivery robots, self-driving transportation systems and many other applications, continue to see very positive development.