Three-stage Junction Detection in Document Images

Engineering drawings contain many curves that intersect at a complex junction. Detection of these junctions is vital for correct recognition of line drawings. In this paper we study the detection of junctions in document images. A threestage junction detection is proposed to detect junctions in document images. The first stage works at the pixel level. Any pixel with more than two neighbors is labeled as low-level-junction. The second stage works at connected component level. All low-level-junctions adjacent to other low-level-junctions are merged into mid-level-junctions. The third stage combines mid-level-junction at a distance less than a threshold into a high-level-junction. Experimental results on real scanned images of mechanical engineering drawings show a precision rate of 82.38% and a recall rate of 97.29%.