Instance-based skew estimation of document images by a combination of variant and invariant
Seiichi Uchida, Shinichiro Omachi, Megumi Sakai, Koichi Kise, Masakazu Iwamura · 2007
A novel technique for estimating geometric deformations is proposed and applied to document skew (i.e., rotation) estimation. The proposed method possesses two novel properties. First, the proposed method estimates the skew angles at individual connected components. Those skew angles are then voted to determine the skew angle of the entire document. Second, the proposed method is based on instancebased learning. Specifically, a rotation variant and a rotation invariant are learned, i.e., stored as instances for each character category, and referred for estimating the skew angle very efficiently. The result of a skew estimation experiment on 55 document images has shown that the skew angles of 54 document images were successfully estimated with errors smaller than 2.0 degree. The extension for estimating perspective deformation is also discussed for the application to camera-based OCR. 1.