Automatic extraction methods of container identity number and registration plates of cars

Three methods of character area localization in noisy images are proposed, constituting the first treatment of automatic recognition of container or wagon identity numbers, or registration plates of cars. The first method, implemented on a two-transputer network, involves a cognitive approach to segmentation, by the search for vertical segments. The second method, based on signal processing, realizes an AR-modeling of the image background, and uses a rupture detection process. The third method is based on the localization of pertinent transitions, i.e. an increase of the gray level crossing over a given percentage of the scale. These methods have been compared on photographs (512*512 pixels on 256 gray levels) of containers and registration plates. The three methods, debugged and checked with the same images, are shown to give good results.<>