Inspection System Automates Glass-Break Analysis

To analyze a piece of glass, a quality-control person places the glass on a flat surface and hits it with a hammer. Then he or she projects the break pattern onto photosensitive blueprint paper, which a camera images (Fig. 1). The person analyzes the blueprint by counting the pieces of glass in the image and evaluating the break pattern according to European Community standards (EU Guideline 92/22).

Measurements take place in two 5x5-cm-sized squares on the blueprint. One square encloses the area with the most pieces, and the other encloses the area with the fewest pieces. For the sample to pass, a square must contain no fewer than 40 pieces nor more than 400 pieces. Moreover, the length of the longest piece cannot exceed 7.5 cm and the area of the largest piece cannot exceed 3 cm². Because people do the inspecting, however, they often get inaccurate counts.

 At first, we thought we could capture an image of the broken glass directly using an electronic camera and then process the image using software. That approach proved impractical because the breaks continue to “grow” and change for several minutes after the hammer's impact. To account for changes in the pattern, we projected the break structure onto a blueprint from which we took the static break pattern using a camera. The blueprint “integrates” all the changes, accumulating them in a single image. We still rely on an inspector's subjective judgment to select the best places for the two 5x5-cm-sized squares.

Our system used a standard Sony XE-75 video camera and a National Instruments PCI-1408 frame-grabber board to acquire images from a blueprint. We wrote the control and image-analysis software with one of our own software tools and “G,” the language used with LabView 4.01. The system automatically recognizes the 5-cm2 squares marked by a quality-control person, and counts and characterizes the pieces of glass according to the European Community standard (Fig. 2).

At the beginning of the project, our colleagues at Sekurit Saint Gobain believed that results from our inspection system should exactly match the number of glass pieces counted by a person. Our first task was to have the inspection system come as close as possible to this number. We found, though, that depending on the person who did the counting, the number of pieces counted could vary greatly. So, we redefined the task to allow for a deviation of 10 counts from the results determined by visual counting.
  

To overcome changes in contrast in the blueprints from test to test—something that humans can quickly adapt to—we adopted fuzzy logic and neural networks that mimic the human brain. These technologies gave our system the capability to learn from experience even when the training samples were difficult to analyze.¹ To implement the fuzzy logic and neural networks for data analysis and control, we used our own software tool, called DataEngine.

When we tested the inspection system, results showed that a fuzzy-clustering algorithm detected the boundary lines of the square area much faster than a Hough-transformation—a standard image-analysis algorithm. To save computer time, we performed the “thinning” of break pattern lines by implementing a multilayer perceptron, a special neural network invented in 1958.² Thinning the lines between the pieces of glass makes them easier to identify, measure, and count.

The list below shows all the image-processing steps the software performs after the system acquires an image:

• Identify the lines that enclose the manually picked region of interest, and isolate this region;
• Compute edge enhancement and combine with gray-level information to determine appropriate threshold;
• Fill holes and smooth boundaries;
• Perform thinning;
• Identify ends of unconnected edges and close suspected gaps in boundaries of objects;
• Count and label objects; and
• Find largest and longest pieces.

The automated inspection system, which took us about six months to develop, is now in use at Sekurit Saint Gobain. It provides accurate results, and it frees people from the tedious and error-prone task of manually counting pieces of broken glass. T&MW

FOR FURTHER READING
1. You can find more information about fuzzy image processing at this Web site: pami.uwaterloo.ca/tizhoosh/fip.htm
2. For more information about perceptrons, refer to, Introduction to Neural Networks, by Jeannette Lawrence, California Scientific Software Press, Nevada City, CA. 1993. 530-478-9040. www.calsci.com.

Manfred Fochem received his degree in electrical engineering in 1992. Since then, he has worked for Management Intelligenter Technologien (Aachen, Germany). He has been working as a project engineer in acoustical and optical quality control. mf@mitgmbh.de.

Peter Wischnewski received his degree in physics in 1995. He now works as a LabView/BridgeView consultant. He has worked on the measurement and data-analysis aspects of acoustical, vibrational, and optical quality control.