Character string extraction from scene images by eliminating non-character elements

This paper discusses a method to extract character strings from scene images. In this method, the Canny edge detector is first applied to a scene image, and the binary edge image is then obtained. Next, small edge elements are separated from large edge elements because edge elements from characters are much smaller than edge elements from non-character objects such as signboards etc. However, many of the small edge elements are noises, that is, most of them are obtained from non-character objects such as roadside trees etc., and so we first remove some of the noises by a dilation operator, a mathematical morphology operator. After that we eliminate the remaining noise edge elements by checking the local segment densities. Finally, a fuzzy inference system is applied to detect character strings from the remaining small edge elements. The performance of the proposed method is examined by computer experiments.