Analysis of manuscript characters by the moments of wavelet packets

The extraction of characteristics is one of the principal components of all form recognition systems. This work relates to the use of wavelet moments as characteristics for the recognition of manuscript characters. We propose here new families and new methods of treatment of this type of characteristic. Encouraging experimental results were obtained: the most discriminating 39 characteristics among 168 others were extracted and selected. These were validated using two classifiers, 1-NN and SVM. The first gives an error rate of 4.56% (leave-one-out), and the second an error rate of 2.99% (hold-out).