Detection of Quadrilateral Document Regions from Digital Photographs

Abstract

A modern smartphone is a convenient and capable tool for document capture. However, without proper enhancement, a casually taken digital photograph is likely to bear visual artifacts of its capture and offers limited utility as a document. Detection of quadrilateral document borders enables perspective correction and cropping, two critical tasks in the enhancement pipeline. This paper presents a method for detecting quadrilateral document borders based on local features extracted from line segments and a machine learning framework. In contrast to many previous methods, this method does not assume printed text content and is thus applicable to a broader range of documents. This method encompasses: 1) shape-based rejection rules that reduce the pool of candidate quadrilaterals; 2) features that take document area, borders, and collinear content into account; 3) a metric that ranks quadrilaterals according to a ground-truth quadrilateral; and 4) a machine learning framework based on learning-to-rank. Experimental results on real-world images show that the listwise learning method ListNet performs significantly better than the pairwise learning method Ranking SVM for our application. The results also show that our multi-feature method significantly outperform an existing single feature method.