DECEMBER 6--Siemens (www.siemens.com) researchers have developed a fast and robust image-recognition sensor that obtains information solely from the color composition of an image. This makes it possible to distinguish between even complex images within fractions of a second. The sensor can thus determine whether a package contains the correct content, for example.
Up until now, sets of data with volumes of two to three megabytes each had to be compared in order to analyze images--a procedure that only expensive processors are fast enough to conduct. As reported in Pictures of the Future, the new color sensor reduces data volume to less than 4 kbytes by means of intelligent algorithms stored in its hardware. This makes it possible to analyze images extremely rapidly.
The sensor consists of a camera chip with a resolution of 640 x 480 pixels, similar to that of the CMOS chip in a cell-phone camera. The system creates a data list after the image has been recorded. This list contains only the color spectrum and the brightness level of the picture. The information can be imagined as something like a three-dimensional cloud made up of points. Each image is therefore depicted through its own unique cloud of points, which the system compares at lightning speed with a reference image.
Among other things, the new system can check product labels on conveyor belts. However, the sensor must first be "taught" to do this. Specifically, it is "shown" the labels it must distinguish between several times until it is able to recognize their color composition regardless of how the labels are arranged. The only thing users have to do is to roughly focus the image sharpness; the flash, comparison of white areas, and brightness adjustments are all carried out automatically. In the checking mode, the sensor differentiates within 30 ms between "correct" and "incorrect" images and displays the result via light signals. The system could conceivably be installed above a conveyor belt that transports bottles in order to check the caps solely on the basis of their color composition, thus ensuring that only one type of beverage is packed into a crate.
The system is expected to be ready for market launch in 2005, and Siemens engineers believe its recognition time could be reduced to 10 ms by combining the camera chip and the image-processing unit--currently two separate components--into one chip.