An Image Based Performance Evaluation Method for Page Dewarping Algorithms Using SIFT Features

Abstract

Dewarping of camera-captured document images is one the important preprocessing steps before feeding them to a document analysis system. Over the last few years, many approaches have been proposed for document image dewarping. Usually optical character recognition (OCR) based and/or feature based approaches are used for the evaluation of dewarping algorithms. OCR based evaluation is a good measure for the performance of a dewarping method on text regions, but it does not measure how well the dewarping algorithm works on the non-text regions like mathematical equations, graphics, or tables. Feature based evaluation methods, on the other hand, do not have this problem, however, they have following limitations: i) a lot of manual assistance is required for ground-truth generation, and ii) evaluation metrics are not sufficient to get meaningful information about dewarping quality. In this paper, we present an image based methodology for the performance evaluation of dewarping algorithms using SIFT features. For ground-truths, our method only requires scanned images of pages which have been captured by a camera. This paper introduces a vectorial performance evaluation score which gives comprehensive information for determining the performance of different dewarping methods. We have tested our performance evaluation methodology on the participating methods of CBDAR 2007 document image dewarping contest and illustrated the correctness of our method.