Methods for invariant signature classification

We present a comparison of several statistical methods to carry out an automatic recognition of signatures. To perform the method we use 6 subjects and the calibration set is composed of 50 signatures for each class. The recognition process consists on the computation of 48 features for each image of the calibration set. A feature extraction process, based in canonical variables analysis, is carried out in order to reduce the number of variables used. Finally, the classification process is performed by using different statistical methods: PCR, PLS, LDA, SIMCA, DASCO, and others. The results obtained show that incorrect signature detection errors were less than 3% in all the techniques considered. However, by using the linear discriminant analysis (LDA) the total error was less than 0.2%. Moreover, the use of LDA is suggested due to the speed of the algorithm. These results prove the utility of this technique for signature automatic recognition