A neural-network-based online signature verification system using vector autoregressive modeling and a novel velocity segmentation scheme

A new online handwritten signature verification system that employs both shape and dynamic features is presented in this work. The main design objective of the system is to obtain a balance of the trade-off between the system's accuracy and practicality. After acquiring the signature sequence using a tablet device, the acquired raw signature sequence is then preprocessed to prepare the sequence for normalization purpose. Next, the preprocessed signature sequence is divided into a number of smaller segments based on a novel scheme developed in this work, which uses the minimum velocity points of the signature sequence as segmentation boundaries. Each of the uniformly spatial-spaced signature segment is treated as a two element vector sequence (xj,yj), and modeled by a multivariate autoregressive (MVAR) model. A one-dimensional autoregressive AR model is also used to model the velocity signal of the signature sequence after smoothing the sequence by applying a Finite Impulse Response FIR low pass filter. The extracted shape features are the MVAR coefficients from the pen position information of each segment; and the extracted dynamic features are the AR coefficients from the smoothed velocity signal of the signature sequence in addition to the total execution time of the signature. The extracted features vector is re-processed by a newly developed Features Consistency Filter, a statistically based mathematical algorithm that weights the extracted features, keeps the most consistent features, and eliminates the inconsistent ones. The filtered extracted features are used together to train a Multi-Layer Perceptron (MLP) Neural Network. A performance evaluation of the system on the development signature database showed system accuracy of 99.8% in a Random Forgery Test (RFT), 98.88% in a Casual Forgery Test (CFT) and 98.63% in a Skilled Forgery Test (SFT). The dissertation also addresses the implementation of a practical system in gums of designing the system to be trained with fewer samples, proposing overall architecture for the practical implementation, and presenting a real time demonstration setup of the system.