On Indexing Handwritten Text

This paper deals with one of the new emerging multimedia data types, namely, handwritten cursive text. The paper presents two indexing methods for searching a collection of cursive handwriting. The first index, called word-level index, treats word as pictogram and uses global features for representing the cursive words and their retrieval. Each word (or stroke) can be described with a set of features and, thus, can be stored as points in the feature space. The Karhunen-Loѐve transform is then used to minimize the number of features used (data dimensionality) and thus the index size. Feature vectors are stored in an R-tree. The second index, called stroke-level index, treats the word as a set of strokes. We implemented both indexes and carried many simulation experiments to measure the effectiveness and the cost of the search algorithm. The proposed indexes achieve substantial saving in the search time over the sequential search. Moreover, the proposed indexes improve the matching rate up to 46% over the sequential search. The word-level index is suitable for large collection of cursive text. The stroke-level index is more accurate than the word-level index, but the stroke-level index is more costly than the word-level index in terms of the search time.