A robust and multiscale document image segmentation for block line/text line structures extraction

This paper presents a new approach dealing with the extraction of blocks of lines and with discrimination between the text line and nontext lines structures. This approach combines a multiresolution textural description of the document image with a cooperation between adjacent resolution levels. At the same time, a not purely bottom-up structural procedure allows us to construct in a very robust way text lines and nontext lines structures and group them into fairly homogeneous blocks. The efficiency of the overall method is demonstrated by its ability to segment and extract destination address block (DAB) on quite diverse and complex images of flat mail pieces