Local and Global Feature Selection for On-line Signature Verification
ICDAR '05 Proceedings of the Eighth International Conference on Document Analysis and Recognition
| Hi-index | 0.00 |
In this paper we approach feature set selection phase in signature verification by applying the method for selection of regression variables based on Mallows Cp criterion for regression. In this way we identify best feature subsets of various sizes for each user of our database on the basis of his/her ten genuine and ten random forgery on-line signatures. Among these subsets we select the best subset that have Cp value closest to p, where p is the number of regression coefficients. Thus, we obtain for each user the best feature subset of a different size. Our aim is to check whether there are common features among best feature subsets for all users which will justify the removal of the rest features from the initial feature set. The results obtained with the database of 140 signatures collected from fourteen users demonstrated that we cannot restrict to common feature set valid for all users but instead of that we have to consider each user best feature set separately in signature verification.