A Clustering-Based Algorithm for Automatic Document Separation

For text, audio, video, and still images, a number of projects have addressed the problem of estimating inter-object similarity and the related problem of finding transition, or ‘segmentation’ points in a stream of objects of the same media type. There has been relatively little work in this area for document images, which are typically text-intensive and contain a mixture of layout, text-based, and image features. Beyond simple partitioning, the problem of clustering related page images is also important, especially for information retrieval problems such as document image searching and browsing. Motivated by this, we describe a model for estimating inter-page similarity in ordered collections of document images, based on a combination of text and layout features. The features are used as input to a discriminative classifier, whose output is used in a constrained clustering criterion. We do a task-based evaluation of our method by applying it the problem of automatic document separation during batch scanning. Using layout and page numbering features, our algorithm achieved a separation accuracy of 95.6% on the test collection.