Writer identification approach by holistic graphometric features using off-line handwritten words

Abstract

The biometric identification is an important topic with applications in different fields. Among the different modalities, based-handwriting biometric is a very useful and extended modality, and the most known one is the signature. The use of handwritten texts is researched presenting a biometric system for identifying writers from their handwritten words. A set of feature-based graphometric information has been extracted from off-line handwritten words to implement an automatic biometric approach. Given the handwritten nature of the information and its great variability, a feature selection based on principal component analysis and neural network classifier has been proposed. A fusion block based on neural networks has been added in order to reduce the effect of the data variability due to an increase and stabilization of the accuracy. A dataset composed of 100 writers have been used for the experiments. A holdout cross-validation was applied and the accuracy reached between 99.80% and 100%