A Robust Approach to Document Skew Detection

Abstract

AbstractThis article presents a script-independent approach for the estimation of skew angle of the offline document image. It follows a certain divide and conquer strategy. A document image is segmented into several vertical strips of smaller and equal width and a simple strategy is applied to identify individual lines in each such vertical segment. Next, an association of horizontally adjacent lines in consecutive vertical strips takes place extracting individual text lines of the whole document image. Let \(P_{ij}\) and \(P_{i(j+1)}\) be the center points of the baselines of two adjacent segments of the i-th text line of the input document. The angle of inclination \(\theta _{ij}\) of the straight line segment \(P_{ij}P_{i(j+1)}\) with the positive direction of x-axis is computed. The average of all \(\theta _{ij}\)’s for the entire document image is used as the estimated skew of the document. This strategy of document image skew estimation has been tested on the DISEC’13 (ICDAR 2013 Document Image Skew Estimation Contest) database and its performance is comparable with the state-of-the-art results in terms of both accuracy and computation time. Additionally, this skew estimation scheme has been used on a wide variety of other documents of different scripts, mixed script documents, degraded quality documents, documents containing texts and doodles, etc. The results of skew correction in every such case have been observed to be satisfactory.KeywordsDocument skewSkew correctionText line segmentationVertical projection profileOptical Character Recognition (OCR)