Fusion of HMM's Likelihood and Viterbi Path for On-line Signature Verification
Fusion of HMM's Likelihood and Viterbi Path for On-line Signature Verification
复制标题
HMM似然与维特比路径的融合用于在线签名验证
DOI：
10.1007/978-3-540-25976-3_29
复制
发表时间：
2004
影响因子：
2.6
通讯作者：
B. Dorizzi
中科院分区：
文献类型：
--
作者：
Van;S. Garcia;B. Dorizzi
我们描述了一种融合来自隐马尔可夫模型（HMM）的两个互补分数的方法，用于在线签名验证。签名是使用数字化仪获取的，该数字化仪捕获笔的位置、笔的压力和笔的倾斜度。当输入签名上获得的两个相似度分数的算术平均值高于阈值时，作者被认为是真实的。第一个分数与对所声称身份的签名进行建模的 HMM 给出的可能性相关；第二个分数与 HMM（维特比算法）对输入签名给出的最可能路径有关。我们的方法在 BIOMET 数据库（来自 87 个人的 1266 个真实签名）以及飞利浦在线签名数据库（来自 51 个人的 1530 个签名）上进行了评估。在 Philips 数据库上，我们研究了训练数据量的影响，在 BIOMET 数据库上，我们研究了时间变异性的影响。进行了多项交叉验证试验以报告可靠的结果。我们首先将我们在 Philips 数据库上的系统与 Dolfing 的系统在他的协议之一（训练 HMM 的 15 个签名）上进行比较。在这些条件下，我们达到了 0.95% 的等错误率 (EER)，而 Dolfing 之前获得的 EER 为 2.2%。当仅考虑 5 个签名来训练 HMM 时，仅依赖于可能性的最佳结果在 BIOMET 数据库上产生 6.45% 的 EER，在 Philips 数据库上产生 4.18% 的 EER。当通过简单的算术平均值融合两个分数时，BIOMET 数据库的错误率降至 2.84%，Philips 数据库的错误率降至 3.54%。
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.