Designing a biometric system, the identification of appropriate features is mostly done based either on expert knowledge or intuition. But there is no guaranty that the extracted features are leading to an optimal authentication performance. In this chapter, statistic methods are proposed to analyse biometric features to select those having a high impact to authentication or hash generation performance and discard those having no or bad impact. Therefore, a short overview on recent related work is given, and appropriate feature-selection strategies are suggested. An exemplary experimental evaluation of the suggested methods is carried out based on two algorithms performing verification as well as biometric hash generation using online handwriting. Test results are determined in terms of equal error rate (EER) to score-verification performance as well as collision reproduction rate to assess hash generation performance. Experimental evaluation shows that the feature subsets based on sequential backward selection provide the best results in 39 out of 80 studied cases in sum for verification and hash generation. In the best case regarding verification, the same feature analysis method leads to a decrease of the EER from 0.07259 down to 0.03286 based on only 26 features out of 131. In hash generation mode, the best results can be determined by only 26 features. Here, the collision reproduction rate decreases from 0.26392 using all 131 features to 0.03142.
Handwriting biometrics - feature-based optimisation, Page 1 of 2
< Previous page Next page > /docserver/preview/fulltext/books/sc/pbse004e/PBSE004E_ch9-1.gif /docserver/preview/fulltext/books/sc/pbse004e/PBSE004E_ch9-2.gif