- This article is filed under:
- » View All
Vision Software Strives For More Functionality in Simpler Packages
by Winn Hardin, Contributing Editor - AIA Posted 03/17/2003
Putting a finger on the most important technological development for machine vision is not that difficult, it’s the modern microchip. But without software to occupy the processor, a microchip alone is just a few dollars worth of metal and plastic. The urgent question today is: What are machine vision (MV) suppliers doing with all those advances in clock speeds, threaded instruction sets and parallel processing capabilities?
‘‘Years ago, we had to do a lot of the image processing in dedicated hardware,’‘ said Bob Kuhl, director of product development for PPT VISION (Eden Prairie, MN). ‘‘A lot of crunching the pixels. Now you’re seeing with the advent of faster CPU’s, there’s more horsepower available and you can do more feature-based implementation where you’re gathering high-level points and looking at the geometry of the points for comparison. And that’s a powerful paradigm for doing things.’‘
Patterns and templates
Older MV systems worked with pixel grayscale values, or pixel-to-pixel comparisons to find features, but changes in illumination, optical aberrations inherent to the imaging system, and other factors made each application unique and difficult to scale. Normalized correlation helped to further the idea of ‘‘turnkey systems,’‘ but newer pattern matching techniques offer the most robust solution to date.
‘‘Illumination, rotation, scale differences and occlusion [difficulties] were the big things that led to pattern matching technologies,’‘ explained Steve Cruickshank, principle product manager for PC based products at Cognex (Natick, MA), and proponent of Cognex’s VisionPro software with PatMax technology. Pattern matching techniques use vectorized feature information rather than indivdual pixel intensity data to establish relationships and shapes. These shapes can be resized, skewed, distorted and repositioned to find features on real world products that resemble, but don’t perfectly match, the Golden Part or perfect product. This approach is directly related to the additional horsepower afforded by more powerful microprocessors.
Variations on the pattern matching technique include the use of Hough transforms to improve edge detection. To take advantage of those and other more powerful image algorithm techniques such as adaptive thresholding that are insensitive to variations in lighting and require less setup time, PPT’s Kuhl said the new IMPACT smart sensor line uses a PowerPC CISC chip to boost processing and create a product category of machine vision micro-system that can run algorithms that used to run only on host-based systems. Hough transforms and enhanced edge detection techniques are proving extremely valuable in reading damaged bar codes, foil packs and other non-planar objects where features are easily distorted. ‘‘Even if a simple object, like an O-ring is squashed, we can still locate it,’‘ said Matrox Imaging’s (Dorval, Quebec, Canada) product line manager, Pierantonio Boriero about the company’s new Edge Finder software tool.
Calibration and color
To further reduce the effect of environmental variants, such as lighting, rotation, etc., companies are making use of more calibration information. For PPT calibration information leads to adaptive thresholding where the system determines environmental IMPACTs on the inspection process and adjusts accordingly. For Cognex’s Carl Gerst, manager for the In-Sight vision sensor product line, enhanced edge detection is pushing his product line deeper into robot calibration where customers need ever more precise spatial measurements.
While calibration has helped these applications, perhaps the greatest influence of enhanced calibration algorithms can be felt in the world of color MV systems. ‘‘What we’ve realized is that our main differentiator is our ability to provide colorimetric images that are very accurate. We incorporate a lot of software up front to accommodate the inadequacies of lighting in this area. We do a lot of calibration to make that work,’‘ explained Bud Patel, director of marketing at Applied Vision (Akron, OH). This calibration step is enabling a new patent-pending feature that will allow the sharing of objective color images between remote locations, Patel added
DVT Sensors (Norcross, GA) has taken a slightly different approach with its SpectroCam. This device combines a grayscale CCD with a prism to separate the incoming wavelengths spatially and make very precise intensity measurements based on wavelength intensity. ‘‘The algorithm looks at the spectrum and compares that to a learned spectrum, so you compare every aspect of the light that’s reflected off that part,’‘ said Mike Schreiber, technical support specialist at DVT.
Programming and Operation
While algorithms – whatever their form and purpose -- are the heart of a MV system, most agree that the best system is one that does not draw attention to the complex computations going on behind the user interface. ‘‘The more robust the algorithm, the easier it is to set up, configure and get the system up and running,’‘ said Matrox’s Boriero. ‘‘The design goal is to minimize operator intervention, but that goes hand in hand with robust algorithms.’‘
The main differentiator between object based set up and text based programming approaches lies within the personality of the installer. Systems like Applied Vision’s Genius systems and Cognex’s In-Sight system are made so people with little or no MV knowledge can set up the system.
Applied Vision represents one end of the spectrum, providing turnkey systems for specific applications. ‘‘From our stand point, if the algorithms are robust, then it doesn’t matter if the guy knows how to set it up all that well,’‘ explained Dr. Richard Sones, Applied Vision’s chief technical officer. Sones has designed their systems with oversized touch screens sporting large buttons to accommodate operators that might wear gloves, and graphics rather than text to set thresholds and other parameters.
Cognex’s Gerst has adopted a spreadsheet approach to setting up the In-Sight vision sensor’s tools. The spreadsheet environment also provides the ability to create a custom operator interface that allows operators on the factory floor to view display results, adjust tolerance settings via control boxes and to manipulate tool graphics extensively during set up. Many of these systems provide networking to increase the number of users capable of accessing the system. This allows access to the system by integrators or application engineers in addition to operators.
Powerful machine vision micro-system, such as PPT’s IMPACT, target integrators. Although they use object based programming approaches, systems like IMPACT also include the flexibility of adjusting the algorithms through proprietary and/or standard programming languages like C++ and Visual Basic. Optional software like IMPACT’s Inspection Builder or Cognex’s Vision Pro QuickStart prototyping module allow integrators to access both high level object based libraries as well as ActiveX and other extensible libraries, enabling integrators to create their own turnkey systems based on their specific industry knowledge while leveraging the economics of a standard, powerful hardware platform.
Advances in microprocessor performance have enabled multilevel approaches to MV systems. These systems have multiple points of entry, making themselves available to advanced programmers as well as manufacturing line managers that have little or no experience in automated vision. As Moore’s law continues to defy physics and continue its march forward, microprocessor enhancements may lead to neural network architectures and kinetic programming methodologies that take automated vision to applications yet unimagined.
There are currently no comments for this article.
Leave a Comment:
All fields are required, but only your name and comment will be visible (email addresses are kept confidential). Comments are moderated and will not appear immediately. Please no link dropping, no keywords or domains as names; do not spam, and please do not advertise.