Wordspotting on Historical Document Images

Abstract. In this dissertation innovative methods of wordspotting on historical printed documents are presented. In particular, two methods based on document segmentation on word level have been developed. The first method uses a hybrid feature scheme for word matching based on zones and projections. It also uses a process of creating query key-word images for any word using synthetic data. The synthetic words are created using images of individual characters taken from the processed documents. The method also presents a process allowing user feedback in order to improve the final results. The second method uses the Dynamic Time Warping (DTW) algorithm for comparing word images. It assist the transition between the synthetic data and real data comparison. Syn-thetic data and real data differ and DTW allows a better alignment be-tween the features of the two images. Again, feedback can be applied to improve the results. Furthermore, a method that uses no segmentation on the document images has been also developed. The method overcomes the problem of incorrect segmentation that affect the final results since it detects query keyword images directly on entire document page im-ages. It also allows for partial matching such as detecting word that are included in larger ones. The evaluation of the aforementioned methods showed satisfactory results presenting better performance against com-petitive methods of wordspotting.