SPIE Membership Get updates from SPIE Newsroom
  • Newsroom Home
  • Astronomy
  • Biomedical Optics & Medical Imaging
  • Defense & Security
  • Electronic Imaging & Signal Processing
  • Illumination & Displays
  • Lasers & Sources
  • Micro/Nano Lithography
  • Nanotechnology
  • Optical Design & Engineering
  • Optoelectronics & Communications
  • Remote Sensing
  • Sensing & Measurement
  • Solar & Alternative Energy
  • Sign up for Newsroom E-Alerts
  • Information for:
    Advertisers
SPIE Photonics West 2019 | Register Today

SPIE Defense + Commercial Sensing 2019 | Call for Papers

2019 SPIE Optics + Photonics | Call for Papers

SPIE Journals OPEN ACCESS

SPIE PRESS

Print PageEmail Page

Defense & Security

Future Vision: A Sneak-peek Into Machine Learning

Jürgen Beyerer of Fraunhofer discusses machine vision and machine learning.

14 August 2018, SPIE Newsroom. DOI: 10.1117/2.2201808.01
Jurgen Beyerer, Fraunhofer

Inspired by one of humans' strongest senses, Dr. Jürgen Beyerer has had a longstanding fascination with machine vision and machine learning. Since vision is a primary sense of human beings, he says, trying to find technical ways to imitate it has always fascinated him, career-wise. "We humans guide ourselves by seeing the world, and therefore, from the beginnings of my studies, that's been particularly compelling for me: to find ways for machines to imitate what human beings do with their eyes and their brains."

In industry, he notes, many tasks and actions such as visual quality inspection or vision-guided assembly, are driven by visual sense. "When we think about automation, almost all the things that human beings accomplish by using their eyes are potential applications of machine vision. In the case of automated inspection, for example, there is still a huge potential for automation offering a lot of benefits with respect to better uniformity in quality control, higher productivity, and cost savings. Another high-profile example is autonomous driving, where machine vision plays an increasingly important role. These are strong driving forces for the success story of machine vision."

Beyerer, who will be giving a plenary talk at SPIE European Security + Defence in September, knows of what he speaks. The managing director of the Fraunhofer Institute of Optronics, System Technologies and Image Exploitation (IOSB) in Germany, he also holds the Chair for Interactive Real-Time Systems (IES) at the Institute of Anthropomatics and Robotics at the Karlsruher Institute for Technology (KIT). While the IES research activities apply methods of automatic visual inspection, pattern recognition, and control systems for monitoring and controlling industrial processes, Beyerer is also currently, in his role at the Fraunhofer's Group for Defense and Security, researching the transfer of data protection procedures from civil security to interactive environments. His plenary topic will discuss the possibilities of coordinated use of advanced imaging systems, target detection, and augmented reality in a battlefield context.

Revolution and renaissance
When he talks about the transformative changes in machine vision and learning in recent years, his voice carries both excitement and awe. In the past, he says, machine vision was engineered into technical systems from a variety of methods and tools. "Things changed with the renaissance of convolutional neural networks (CNNs) about ten years ago," says Beyerer. "A revolution started: with huge amounts of data, progress in learning algorithms, availability of high-performance computing, especially due to GPUs, and, with CNNs, new, deep architectures, it became possible - with little or nearly no engineering effort - to build and train black-box vision systems that outperformed well-engineered systems significantly. It's been fascinating to develop creative image acquisition setups, to make things accessible that human eyes cannot see - invisible wavelengths, very fast processes, etc. - and to transfer many different concepts from mathematics and informatics in order to exploit image data. A CNN for image processing has a very simple structure," he continues, "but there are many connections in it, millions of parameters. It is a universal mathematical approximator for functions, but a very mighty one. You can learn about it, without understanding what's happening inside. So for you it's a black box; but these black boxes perform better than the sophisticated, engineered systems. That was a big surprise for all of us who work in the field of machine vision and image processing."

And, Beyerer points out, the potential of deep learning when it comes to machine vision, is far from fully explored or exploited. "In the future," he says, "deep learning, especially of CNNs in different creative versions, will be more and more important for machine vision. All our research groups concerned with machine vision are also involved machine-learning research, working to explore its great potential."

Importance of system integration
At Fraunhofer, Beyerer contributes to an institutional research model built to generate scientific excellence and successful applications. With public funding for the institutes at about 30%, the rest of the funding comes via competitive acquisition of research and development projects, with more than a third ideally covered by projects from the industry and economy sectors. "Along with its mission to apply research results successfully outside the institutes, Fraunhofer has developed its own specific culture to focus on excellent and effective research and efficiently developed, deployable solutions," he explains. "We have excellent scientific and technological capabilities, and additionally, we know the application branches and we know their demands. Two things have to come together: on the one hand you have to be good in your competencies, and on the other hand you have to understand the market, you have to speak the language of the user, the user of a field of application. We do a lot of machine vision of automated visual actions in the field of glass, for example, glass for windscreen or for illuminations or optics and so on, which are very difficult to inspect. They have standards, their own language, their own requirements for things that can be done or can't be done in their companies, and you have to know it all. It takes a lot of time to really understand all of the finer details. It is very important not only to be good in our own competencies - image processing optics, computer science, and so on - but also to deeply understand what's going on in the industry and the way they want to apply your science."

The most exciting research, he finds, comes from more than one direction. "On the one hand, it is important for the development of machine vision to approach your research by trying to solve difficult practical problems; difficult problems inspire - that's clear - and this is a kind of market pull: the market comes, says, ‘We have a difficult problem, we don't know how to solve it. Please show us how we can do it,' and our engineers and our scientists look for a clever solution. On the other hand, it is also very important to pursue interesting new ideas and concepts to find out, what their potential could be with regards to machine vision. This is a kind of technology push. For good and fruitful research, both are required, both are necessary."

System integration, he notes, is also critical for successful problem-solving, something he has learned both as a researcher and while working on vision systems in industry: "Clever algorithms alone cannot solve real world problems. According to our experience, for challenging machine-vision tasks, the actual algorithms for image processing contribute only about 10% to the solution of a problem at hand. Machine-vision systems require and comprise much more in terms of mechanical, optical, electrical, and software engineering, as well as a good understanding of the real-world problem to be solved - and of the people who will use the system."

Beyond the technical framework, his advice is just as dynamic and engaging as the monitoring technologies that encompass Beyerer's professional specialty: "Read a lot!" he exclaims. "Be curious! Be pragmatic! Stay open-minded to new ideas, preserve enough creative space to consider even unorthodox ideas and" - my own particular favorite - "discuss your thoughts with young people."