An Algorithm for Reducing Text Line Candidates of Incorrect Orientation

Japanese documents often contain both horizontally and vertically printed text lines in the same page. It has been required for document analysis systems to detect correct orientation of text lines and to select text line candidates of correct orientation. We designed an efficient framework for the procedure and developed some algorithms which reduce text line candidates of incorrect orientation. However, our previous system could handle only text lines with slight skew (f 10"). In order to improve the performance of our document analysis system, we recently developed an improved algorithm for text line extraction which can handle curved or wavy text lines as well as straight text lines of arbitrary orientation. We have also developed a new candidate reduction algorithm for arbitrarily oriented text lines. This paper describes mainly the detail of the candidate reduction algorithm. The overview of the system is also mentioned. The candidate reduction algorithm is based on the a priori knowledge that inter-line spacing is much wider than inter-character spacing in most documents. Experimental results show that the algorithm works very well for many documents, including very complicated ones, written in both Japanese and English, and also for the documents which contain curved or wavy text lines.