Skew angle detection using texture direction analysis.pdf

Skew angle detection using texture direction analysis Jaakko Sauvola and Matti Pietik?inen Computer Laboratory Department of Electrical Engineering University of Oulu FIN-90570 OULU, Finland e-mail: jjs@ee.oulu.fi Abstract In a document analysis system the skew error detection is one of the most cru- cial parts. This paper describes a method that takes advantage of texture orien- tation analysis in order to find the skew error from the source document. The source image is first shrunk to melt the characters in together, reduce noise and decrease the amount of computing. After pre-processing a texture direction analysis method is applied. This procedure produces an unambiguous angle in degrees as the skew angle. 1. Introduction Skew detection and correction is a very important problem when an OCR algorithm tries to detect and interpret characters from the source document. Many different solutions have been developed. The stage in which the skew detection takes place also varies from the first to the very last one before the OCR phase. The document structure analysis and character recognition are usually done in several phases: scanning and thresholding, image enhancement, skew detection and correction, segmentation, classification and character recognition. Each step must be completed well enough for the performance of the sequence and result to be successful. Steps that follow the skew correction are inefficient if the correction fails. Skew correction may depend on many different approaches. Methods based on Hough transform and run-length smearing [1], cross-correlation [2] and projection histogram [3] have been proposed, as well as methods that rely on white space versus black areas (text/ picture) and their placing. Most of the previous methods are strongly dependent on the amount of textual coverage on the page. Their efficiency usually drops if the text content in page diminishes. This paper proposes a new approach based on texture orient