Skew detection and reconstruction based on maximization of variance of transition-counts

摘要

The input document images with skew can be a serious problem in the optical character recognition system. A robust method is proposed in this paper for skew detection and reconstruction in document images which can contain less text areas, high noises, tables, figures, flow-chart, and photos. The basic idea of our approach is the maximization of variance of transition counts for the skew detection and text-orientation determination. Once the skew angle is determined, the scanning-line model is applied to reconstruct the skew images. 103 documents with great varieties have been tested and successfully processed. The average detection time of A4 size image is 4.86 s and the reconstruction time is 5.52 s. The proposed approach is also compared with the existing algorithms published in the literature and our method gets some significant improvements in skew detection and reconstruction.