A Vector Quantization Based Feature Descriptor For Online Signature Verification

This work proposes a scheme to authenticate the veracity of an individual through his/her online handwritten signature. The main contribution is in deriving a set of descriptors for verification based on a pre-generated codebook. The codebook, as such, comprises a set of codevectors that are obtained from a Vector Quantization based scheme applied on feature vectors of enrolled signatures of the user in question. The descriptors take into consideration, the score of each of the attributes in a feature vector, that are computed with regards of the proximity to their corresponding value in the assigned codevector. A second contribution of the paper deals with the idea of matching the signatures by associating a consistency factor to the descriptor of each of the codevectors. The consistency factors are pre-learnt by using the set of reference signatures enrolled to the system. In addition, we empirically demonstrate that the traditional dynamic time warping system used in conjunction to that built from the codebook descriptors can help improve the error rates. Experiments conducted on the MCYT-100 echo the efficacy of our proposal.