On-line cursive letter recognition using sequences of local minima/maxima

This report presents the design and implementation of an on-line cursive letter recognizer using sequences of local extrema. Recognition tests were performed on single writer's data. The results for top five alternatives exceeded 90% (Table 4). The matching is performed in three stages: • letter size normalisation (Figure 1); • limiting the recognition domain using sequences of local extrema and inexact matching (Figure 3) with a database; • evaluating the goodness of match for each element of the obtained limited recognition domain. Two methods of letter size normalisation were considered and tested: • uniform (Figure 1b); • non-uniform (Figure 1c). The uniform normalisation approach proved to be better (Figures 5, 6 and 7). Three different criteria are used for the last stage, resulting in three different " little recognizers " : • position of the local extrema (Figure 4a); • characteristic loci (Figure 4b); • loci angles (Figure 4c). The results of the three recognizers could be combined together in order to improve the recognition rate. The recognizer needs to be trained. The databases are built in two main stages: • creating the database of sequences of local extrema; • deriving detailed feature data for ranking the alternatives. The first stage is currently manual and may not lead to the best local extrema sequence database. The second stage can still need some supervision, depending on the specific extrema sequence database. The recognizer possesses a potential for incremental learning. At run time both new sequences of local extrema can be added, and the detailed features used for the ranking can be adjusted to better reflect a particular writing style. Further work is necessary: • tests with other writers' data; • conversion of the difference between patterns into the confidences of the match; • combination of results of different classifiers; • integration into the on-line cursive script recognition system, testing; • comparison with the existing letter recognizer using vector direction encoding.