exit bar




Electrical and Computer Engineering
& Industrial/Economic Development

 Spring 1996

"Within five to 10 years, we'd like to develop generic systems that can serve broad reaching applications...which would lead to significant improvements in efficiency and quality worldwide.


"Right now, we're pushing the envelope of the technology with machine vision applications - and that's where university research belongs."

-Richard Conners
Spatial Data Analysis Laboratory


If Machines Could See..

Heavy industry wants to develop computer systems that can "see" - and they are turning to Tech computer vision experts for help.

The lure of improved efficiency, accuracy and quality is leading many manufacturers to seek out computer systems that can visually inspect materials and process the data in real time. Machine vision systems can look for defects in materials, inspect part tolerances, monitor inventories, or even guide robots through hazardous environments.

Machine vision is well established in the highly automated electronics industry, where parts are small and tolerances tight. However, vision systems are less common in heavy industrial applications. The larger parts and materials have proved to be a challenge due to the limits of computer speed and the very large size of the image files.

Recently, several different firms, including cabinet manufacturers, and large equipment manufacturers have turned to Virginia Tech machine vision experts for help in developing vision systems.

Identifying cabinet partsphoto of cabinet vision

On one project, Professor Lynn Abbott COLOR="#000000"> and a team of graduate and undergraduate students developed an automated computer vision system that classifies kitchen cabinet components by size, shape, species and color.

The system was developed for American Woodmark Corporation of Winchester, Virginia, the country's fourth largest manufacturer of kitchen cabinets. The facility where the system was installed produces more than 9,000 distinct items - doors, frames and other components - on three modern assembly lines, and then ships the components to other locations for final assembly.

"The system has been very successful," according to Alfred Foster, an industrial engineer at American Woodmark. "We have installed two stages of the system and are identifying cabinet parts by size and shape. We've had the system running two shifts every day since the first of February, and it's fantastic. The accuracy has been 98 to 99 percent. The last I checked, there had been only one misidentification in 20,000 doors."

The system helps the facility to maintain a tighter inventory, and thus improve profitability. "Before the system was installed, if we wanted to know what was in stock, we had to go and do visual counts," Foster explained. "Now that information is on the computer and can be accessed at any time," he added.

Comparing parts to specsphoto of machine vision

Another recent computer vision project was the development of a prototype system that inspects flat parts - sheet metal, for example - and compares them with CAD descriptions for quality control. The project was funded by Hitex Corporation of Lynchburg, Virginia, a manufacturer of specialty textile handling equipment.

When manufacturing large textile handling equipment, Hitex workers must visually inspect the laser cuts on sheet metal. There are very tight tolerances that require the use of calipers and rulers when inspecting the holes. This manual inspection is time intensive. Owner Fred Grzy­bowski was interested in a fast, automatic and accurate system to replace manual inspection.

A team of Virginia Tech researchers worked closely with Hitex to develop and build a small prototype system that is fast enough to accommodate the very large metal pieces that Hitex produces. The system uses a large light box to backlight parts as they are scanned by a moving camera. A PC records the high contrast images; then the information is compared to the stored CAD description. The prototype is currently being evaluated by Hitex.

"There is quite a market for a machine like this. We hope to be able to commercialize it in the future," said Grzybowski.

Limitations and promises

"Now that computers are faster, cheaper, and can handle large image files, we can help heavy industry - those firms dealing with steel, metal, textiles and wood," said Professor Richard Conners, director of the University's Spatial Data Analysis Laboratory, where much of the machine vision work is done. "The real difficulty is that each system must be custom designed and involves unique algorithms. This means a large initial investment. However, with these systems, many businesses can increase their global competitiveness, improve their bottom line and stay in business.

"Within five to 10 years, we'd like to develop generic systems that can serve broad- reaching applications. Then we would be able to increase the number of facilities that can afford machine vision, which would lead to significant improvements in efficiency and quality worldwide," he said.

"Right now, we're pushing the envelope of the technology with machine vision applications - and that's where university research belongs," he added.

MORRPHing ImagesMorrph photo

Two of the challenges facing researchers and developers of computer vision systems are the speed of the computers and the need for custom-built systems and parts. A Department research team has developed a Modular and Reprogrammable Real-Time Processing Hardware architecture that helps to overcome these limitations.

Called the MORRPH board, the system performs very fast, real-time image processing, and provides a user-configurable, modular architecture for cost-effective customization.

The MORRPH board is based on field programmable gate arrays (FPGA), which are gate-level logic resource chips that allow users to create custom configurations, and make changes as needed. The MORRPH board allows up to six FPGAs, along with additional hardware resources.

"MORRPH is a modular solution in a two-dimensional mesh architecture," explained Tom Drayer (G), who developed the concept while serving as a Bradley Fellow. "What's different about this solution is the ability to embed additional hardware resources into the architecture, closely coupled with each FPGA. For example, if I want to add a memory chip to one FPGA and a hardware multiply chip to another, I can do that.

"You integrate only the parts you need on a board, so it's a cost-effective solution," he added. He described one application where the system was used for the real-time scanning of hardwood lumber to determine defects in the lumber. The system integrated X-ray scanning, laser scanning and color imagery, and processed the data from all three on a MORRPH board.

Because of its horsepower and ease of use, the MORRPH board is expected to have broad application where image processing speed is critical.

Various research and development teams have been established by the Spatial Data Analysis Laboratory for applications development of the board. Several graduate students are now working on software development and debugging tools, and 10 undergraduate students comprise the MORRPH design team exploring application development.

The development team is pursuing a patent, and a start-up firm, Pixell, Inc., has been created to commercialize the system. In addition to Drayer, the original development team included Will King, a master's student who developed the prototype, and Professors Richard Conners and Joe Tront. The prototype was created with funding from the U.S. Forest Service.

The Bradley Department of Electrical Engineering
Virginia Tech

Last Updated, June 10, 1997
Questions or comments about the content: eqb@rightwordonline.com
Technical questions or comments: webmaster@birch.ee.vt.edu