Page Tilt Detection And Layout Analysis Algorithm

Document analysis and understanding is to research into theories and technologies during the whole procedure converting paper documents into electronic formats. Document analysis and understanding is important part of OCR (Optical Character Recognition) system. Prolegomenon introduces briefly OCR system, expatiates status and action of this technique in OCR system and reviews the evolutional course and status quo. Preprocessing is important stage of document analysis and understanding. This paper enumerates several familiar approaches of image smoothing and binary. Based on the experiments, this paper compared performance of these algorithms. Page skew detection and correction is also an impassable part of preprocessing. An innovative approach is presented to detect skew angle, skew detection based on image logic operation and LMS. This paper defines serial logic arithmetic operators, proves properties of these operators, and analyzes the directive action of these properties to skew detection algorithm. Firstly, slantwise page is binaried, then, used logical operator T to act the binary image. We can get a solid graphic block. This block can maximize approximate beeline information of slantwise page boundary. Secondly, we use 4-orientation chain codes to represent four boundary of the block. Finally, we extract approximate beeline information from four chains, and calculates skew angle of beeline based on the LMS. Experiment shows this algorithm is fast and accurate. So this approach has certain theoretic value and practical value. Projection-based recursive algorithm is a classical algorithm for document segmentation. But this algorithm performance is restricted seriously by threshold. Ill-suited threshold will result in super-segmentation of text region. In order to overcome this defect, this paper smartly uses the idea of region combination of Bottom-Up strategy for reference, which improves the performance of traditional algorithm. Region recognition approach is based on connected component features. In addtion, this paper analyzes experiment performances of document segmentation and region recognition in detail and lists respective strong points and defects. As for defects, the author discusses reasons bringing on these defects. In addition, large numbers of figures are showed to illuminate farther. In the end, the paper summarizes harvests and deficiencies of this academic dissertation and presents the opinions about this field and my experiences.