Extraction of character string region by a correlation method

This article describes a three-level thresholding method that is used for the preprocessing for searching character-string regions on a natural image, and a method of defining the region by using the correlation process of two kinds of binary images obtainable from the three-level thresholding. Three-level thresholding is a method by which a high-frequency image is obtained by subtracting a moving-average image from the original image and the high-frequency image is changed to an image consisting of three kinds of elements {-1, 0, 1} using the variance of the image. This method is effective when the brightness of characters and their background are not known or fixed at. When the length of a moving average length is sufficiently longer than the width of a character, the three-level thresholding produces an image in which character regions are represented by 1 or -1. The correlation procedure that is made by multiplying pixel values of the two images, one of which is shifted by the width of a line segment of a character, can generate a character region in high probability. A character region is detected by matching of the correlation image with a rectangular template representing the numberplate. The proposed method has been successfully applied to detection of vehicle registration numbers by a roadside camera.