A Method for Character String Extraction Using Local and Global Segment Densities

摘要

We propose a new method for extracting character strings which have various directions and sizes. In our method, we regard a character string as a rectangular region crowded with short segments. Then, two kinds of segment densities (local segment density and global segment density) are introduced to extract the rectangular regions (i.e. strings). First, regions locally crowded with segments in the input image are extracted as string candidate regions. Next, segments within the extracted string candidate regions are taken out. Finally, rectangular regions crowded with some of these segments are extracted as strings using improved Hough transform. The method was applied to 28 images which involve 276 strings in various directions and sizes. As a result, 91.3/100 of 276 strings were correctly extracted.