Selection of reference signatures for automatic signature verification

Automatic Signature Verification: The State of the Art

“…For instance, contour-based features can be extracted at global level (i.e., envelopes extracted at the level of the whole signature) or at local level (i.e., at the level of each connected component). function is conveyed directly by the acquisition device whereas velocity and acceleration functions can be provided by both the acquisition device [121], [178] and numerically derived from position [40], [51], [346]. In recent years, pressure and force functions have been used frequently and specific devices have been developed to capture these functions directly during the signing process [46], [219], [235], [237], [300], [374].…”

“…Stroke-based signature verification can lead to lower error rates compared to global approaches, since a large amount of personal information is conveyed in specific parts of the signature and cannot be detected when the signature is viewed as a whole [8], [21], [54]- [57], [59], [126], [164], [278], [298]. Furthermore, the verification at stroke level can be performed by DTW [41], [50], [51], [55], also considering multiple function features for stroke representation (like position, velocity, and acceleration) in order to verify both the shape and dynamics of each part of the signature [18].…”

“…When static signature verification is considered, the validity of the reference model has been evaluated according to specific quality criteria, as for instance, intraclass variability that should be as low as possible [3], [78], [79]. In dynamic signature verification, the selection of the best subset of reference signatures has been performed on the basis of the analysis of variance within samples [112] or by considering the stability regions in the signatures, determined by a well-defined analysis of local stability [40], [51]. The selection of the best subset of reference signatures can be avoided at the cost of using multiple models for signature verification [148], [197], [198].…”

“…Several decision combination schemes have been implemented, ranging from majority voting [4], [50], [51], [59], [274] to Borda count [12], from simple and weighted averaging [18] to Dempster-Shafer evidence theory [12] and NNs [15], [17], [26]. The boosting algorithm has been used to train and integrate different classifiers, for both verification of online [122] and offline [329] signatures.…”

“…Since an ME verification system should combine decisions from complementary signature verifiers, sets of verifiers based on global and local strategies [92], [95] and feature sets [123], [125], parameter features and function features [260], static and dynamic features [50], [51] have been used. Several decision combination schemes have been implemented, ranging from majority voting [4], [50], [51], [59], [274] to Borda count [12], from simple and weighted averaging [18] to Dempster-Shafer evidence theory [12] and NNs [15], [17], [26].…”