OCTOBER 22--A new optical technique to detect dents and scratches in car bodies prior to painting has been developed by Germany-based start-up INB Vision, a working group of the University of Magdeburg that commercializes the university's technologies. Said to be more accurate than conventional optical techniques, the method detects dents just 20 μm deep in highly curved surfaces.
Said INB Vision researcher Peter Albrecht, "The concept is similar to human vision, [but rather than using images from the eyes], we use images from the cameras to calculate three-dimensional information about the surface." Conventional automatic techniques often fail to detect defects on complex surfaces, such as the handle-hollow of a car door. Alternatively, edges and corners can be classed as defects. Albrecht and his colleagues claim to have overcome these problems by combining high-resolution photogrammetry with a memory neural network.
Two digital cameras are used to take pairs of images of the sample's surface. A projector, mounted between the cameras, shines a series of gray-scale pixel patterns across the surface. By correlating the sequences of gray-scale pixels seen by each camera, the method obtains the high-resolution, three-dimensional data it needs to detect defects.
To analyze the images for defects, the researchers created an artificial neural network with a memory of the "master" sample or work piece. In its "training phase," the network measures and stores the image data of several master work pieces. This allows the network to later pinpoint any differences between a sample's surface and the surface of its master.
"The key part of this innovation is the memory [of the neural network] and its use in this application," said Albrecht. "The combination of the optical and neural network techniques is new, and our results are excellent."
Using this technique, the researchers can examine an area of 60 cm� in 2 min but are aiming to reduce this time to just 10 s. Albrecht believes that the technique can be used by car-body manufacturers worldwide. "Our next step is to install a test unit in a factory with an operational partner who can commercialize the technology for the automotive industry," he said. "We expect this to happen during 2002."