Page skew is one common type of scanning variation in document image analysis. Often, page skew is first corrected before any follow-up tasks such as optical character recognition (OCR) are carried out. Existing methods of skew correction include the Hough transform class and the projection profile class.

In this work, the authors make progress on the projection profile-based methods and exploit the fact that vertical projection profiles (VPPs) consistently complement horizontal projection profiles (HPPs). Specifically, the authors first compute VPPs making use of the border lines introduced during scanning. Then they compute HPPs and employ a selection procedure, which is essentially an optimization process, to determine which skew is most likely. Experimental evaluation covers several large datasets of different scripts and layout styles.

Jumping out of the micro perspective, it is worth discussing whether there are better ways of handling such scanning variations than in the pre-processing module. As many of the pre-processing techniques show, image transforms on a 2D grid are always approximate, meaning that distortion of the bitmap is inevitable: for example, skew correction. And in many cases, such bitmap modifications are discarded in the processing pipeline. The next question is, after we detect the page skew, is it possible for the follow-up feature extraction and/or classifier modules to compensate for such variations? Bitmap integrity can be valuable.