Rotation of the binary image of a document page for correcting the skew in the case of OCR or signature verification systems entails disfigurement of the shape of the characters, which brings down the performance of the pattern classifier. An elegant algorithm is presented in this paper for distortion-free rotation. Multirate signal processing principles are applied to solve this problem, which was hitherto handled using only heuristics. The resolution of the image is increased during rotation. The rotated image is decimated and thresholded after low pass filtering. Polyphase implementation of the filter is performed for efficient computation. The results show that it solves the problem of mutilations and occasional breaks obtained in simple rotation schemes and further, performs better than the scheme involving interpolation in the original or the rotated domains. Consequent on the better retention of their original shape, the recognition accuracy of originally skewed documents impro...
Kaushik Mahata, A. G. Ramakrishnan