Correcting for Variable Skew

Many techniques have been developed for detecting the dominant skew angle of printed documents [2]. However sometimes more than one skew angle is represented in the document image. Baird [3] points out that the paste-up of text columns may result in slightly (but significantly) different skew angles and notes the importance of segmenting the columns first in order that individual skew corrections can be applied to the individual columns. Spitz [6] and Antonacopoulos [1] have described techniques for detecting the presence of multiple discrete skew angles in a single document image. Spitz showed the angular distribution of the “power” (as defined by Baird) for both instances of multiple discrete skews. This is reproduced in Fig. 1. This paper presents a solution to a different problem, that of documents which contain a continuously variable skew as a function of vertical distance along the page. In Section 2, we describe the type and range of skew angle variability seen. Section 3 describes the segmentation of the image and determination of the dominant skew angle for the individual segments. Section 4 describes the correction of the skew for each of these segments. We discuss the re-assembly of the image in Section 5. Throughout Sections 2 though 5 we will describe the algorithms using a much simplified page image specifically synthesized for this study because doing so makes the processes much easier to understand. In Section 6 we will provide a real-world example.