Detecting Near-Duplicate Document Images using Interest Point Matching

We present an approach to detecting near-duplicate document images using SIFT interest point matching. Given a set of document images, a database is constructed from the SIFT features extracted from each image, stored as a kd-tree. The near-duplicates of a query image are estimated by directly matching its SIFT descriptors with the feature database. We demonstrate the approach on a challenging set of unconstrained Arabic hand and machine written images obtained from the field, consisting of 16,000+ documents. Our experiments indicate that the approach detects near-duplicates with low false alarm rate and outperforms bag-of-words based approach.