Studies In Algorithms For Page Segmentation And Classification Of Document Images

Owing to the fact that the electronic file has its superiorities in efficient storage, retrieval and transmission, it is of great values to transform the traditional paper document to its electronic version. Recently, the research of the techniques for such a transformation became an important area in domain of pattern recognition. Usually, the task of such a transformation is composed of two steps: the first one is to scan the paper document into electronic images while the second is to turn the images into text via OCR(Optical Character Recognition)system. However, many documents contain non-text regions such as figures and pictures in addition to text. Thus, it's necessary to separate non-text areas from text areas so that only pure text regions are input to OCR system. This procedure can be the post-processing of the scan step or the pre-processing of OCR system.The main purpose of this thesis is to study algorithms for page segmentation and classification of document images. In order to make our algorithm proposed in this thesis valid for the skewed document images, a skew detection algorithm based on the Hit-or-Miss operation and the Hough transform was firstly proposed for skew angle detection as well as the skew correction for document images. It has been proved in our experiment that such a skew detection algorithm is superior to other algorithms proposed in the literatures in its speed and accuracy. For the de-skewed document images, an algorithm based on the recursive cuts of projection profiles was proposed for the segmentation and an initial classification of the document images. By introducing a middle point cut process to the traditional projection profile cut algorithm, thesegmentation algorithm proposed here can process document images -with non-rectangular graph regions involved in them. To classify different segmented regions as text regions or non-text regions, two features of the black to white pixel ratio (BWR) and the cross-count (CC) of pixels were used for the classifications. Experimental results have shown that the algorithms proposed in this thesis are robust to noise and have high speed.