Text localization and character segmentation algorithms for automatic recognition of slab identification numbers

Abstract

This paper describes application-oriented text localization and character segmentation algorithms in images. The target text in our application includes many unclear characters due to poor environment as well as the fact that their positions are variable in the images. Consequently, it is difficult to expect a high success rate when using existing text localization algorithms that have been developed for generic texts. Therefore, it is necessary to develop a new text localization algorithm. We propose (1) a coarse algorithm for detecting top and bottom boundaries, (2) a fitness function that is used to decide the true text among the text candidates, (3) two kinds of presegmentation algorithms for calculating the fitness function, and (4) a blank-detecting algorithm that determines whether the text is upside down or not. By the proposed algorithms, input upside-down text is rotated automatically without using any supervised or unsupervised learning methods; further, character segmentation can be done in the process of selecting the true text. To evaluate the algorithms, image data captured by the installed recognition system at Pohang Steel Company (POSCO) are used, and experimental results show that the proposed algorithms are fast and reliable.