Weighted Hough transform on a gridded image plane

Abstract

Detection of skews in input images is an important preprocess for analyzing document images. We propose a novel skew detection method based on vertical as well as horizontal splitting of an image. The center of gravity of the black pixels in each section is used as a sample point for performing the Hough transform, which gives the skew angle of the document image. We extend the Hough transform to incorporate the notion of weight, and apply it to reducing spurious line detection caused by the presence of large non text areas with many black pixels. The weighting function is defined according to the proportion of black pixels in the area, giving a low weight to an area with an extremely high proportion of black pixels. We conducted skew detection experiments on several document images. By using the Hough transform on connected components our method works four times faster than the conventional method and it has no restrictions on the detection range within which skew angles can be detected.