Accepted Papers


Yangli Wang and Rui Song, School of Telecommunications Engineering, Xidian University, Xi’an China


In the capturing process of document images, skew is usually unavoidable. This skew may seriously degrade the performance of subsequent processing steps such as layout analysis and segmentation. As such, skew detection and correction is a necessary step in the document image processing pipeline. In this paper we propose a new skew detection method using character orientation estimation. The first step of the approach is classifying the connected components in a binarized document image into text/non-text. This is accomplished by a convolutional neural network (CNN). The next step of the proposed approach is estimating character orientation using the minimum area rect surrounding a character. Initial document skew is then estimated with the first ?? largest characters and their orientation. This initial skew is then refined by clustering deskewed nearest characters and minimizing sum of the area of their surrounding rects in a small search range. As text/non-text are separated first, and only text is involved in document skew estimation, the proposed method overcomes the drawbacks of projection profile based methods, which can not deal well with documents containing other elements, and are weak in handling multicolumn documents when the lines of different columns are not well aligned. The proposed approach also has the advantage over those methods estimating text line orientation, such as the Hough transform based ones, they have difficulties in choosing representing points for characters as characters are of different height. Simulation results on test document images show that, the proposed approach has achieved a mean estimation error of as low as 0.1169°. It also has the advantage of potentially saving the geometric layout analysis step in the conventional document image processing pipeline.


Character Orientation, Convolutional Neural Network, Minimum Area Rect, Character Cluster, Document Skew Detection