Fusion of HMM's likelihood and Viterbi path for on-line signature verification

We describe a method fusing two complementary scores descended from a Hidden Markov Model (HMM) for on-line signature verification. The signatures are acquired using a digitizer that captures pen-position, pen-pressure, and pen-inclination. A writer is considered as being authentic when the arithmetic mean of two similarity scores obtained on an input signature is higher than a threshold. The first score is related to the likelihood given by a HMM modeling the signatures of the claimed identity; the second score is related to the most likely path given by such HMM (Viterbi algorithm) on the input signature. Our approach was evaluated on the BIOMET database (1266 genuine signatures from 87 individuals), as well as on the Philips on-line signature database (1530 signatures from 51 individuals). On the Philips database, we study the influence of the amount of training data, and on the BIOMET database, that of time variability. Several Cross-Validation trials are performed to report robust results. We first compare our system on the Philips database to Dolfing's system, on one of his protocols (15 signatures to train the HMM). We reach in these conditions an Equal Error Rate (EER) of 0.95%, compared to an EER of 2.2% previously obtained by Dolfing. When considering only 5 signatures to train the HMM, the best results relying only on the likelihood yield an EER of 6.45% on the BIOMET database, and of 4.18% on the Philips database. The error rates drop to 2.84% on the BIOMET database, and to 3.54% on the Philips database, when fusing both scores by a simple arithmetic mean.