We will develop a framework for associating measurements of document image characteristics with image transformation techniques and evaluating the utility of both using a supervised machine learning component of the framework. From a training set of images described in terms of their characteristic values, each with its best transformation method, this component will learn a function that maps combinations of feature values to single best image transformation methods. This research extends the ideas in work by McNamara, Casey, Smith, and Bradburn (1996) and by Cannon, Hochberg, and Kelly (1999). In Phase I, we will develop an application that includes components for: calculating image feature values or quality measures, applying image transformation methods, training the system to associate profiles of image characteristics with image transformation methods, selecting image transformation methods on the basis of this training, evaluating OCR results against available ground truth, and evaluating human readability. Because this phase is centered on development of a complete framework, the initial set of characteristics will have a single measure, and likewise the initial set of transformation techniques will include one method (other than keeping the original image). In the Phase I Option, we will add an additional image characteristic and transformation technique. Anticipated benefits of this research and development will be realized in commercial, non-profit, and public sector markets where there is a requirement for evaluation, capture, and/or retrieval of document contents by any combination of automatic (OCR and full-text indexing) and manual (review, manual indexing, redaction, and retrieval) methods. Applications include high-volume processing of business documents - invoices, proof-of-delivery documents, time sheets etc. - found in virtually every commercial, non-profit, and government organization; archiving and retrieval of historical material; processing of large collections for FOIA and Executive Order declassification; and evaluation of documents acquired during intelligence gathering. In each case, enhanced document images leads directly to increased productivity and accuracy of business processes and higher competitive value to end users who are evaluating, retrieving, and using document images and the content extracted from them.
