An Unconstrained Rotation Invariant Approach for Document Skew Estimation and Correction

The OCR technology is gaining more and more importance in the digitalization of the documents, this is because of its functionality to convert the text data in the image to the machine-encoded text and this machine-encoded text can be further used for processing. The orientation of the digitized document is important for the OCR to recognize the data in the document veraciously. Sometimes due to manual error, the scanned document may not be properly oriented, this condition is called skew of an image. Deskewing is a procedure to align the image properly, before further processing the data in the image. There are many existing approaches for deskewing the image such as mathematical morphology, principal of connected components, projection profile technique, Fourier transform, Hough transform, Radon transform and KL Transform. These methods for deskewing have their own constraints with respect to font style, font size and are not rotation invariant. In this paper, we propose a method which can deskew an image with any degree of skewness using warp-affine transform, Hough transform and feedback of the OCR output. The warp-affine transform is used for adjusting the shape of the background image, Hough transform is used for checking the vertical symmetry of the text and feedback from OCR is used for checking the skewness of 180° and flipped document cases. The proposed method was evaluated on 40 images with the various skew angle and the performance was comparable with the existing techniques in the literature.