Efficient curve-sensitive features for offline signature verification

Abstract

A novel system that is based on the efficient detection of curved lines within an input image is proposed for the purpose of offline signature verification. This is achieved by first trans­forming the signature image into five different representations of said image in polar coordinates, where each polar representation originates at a different strategic point. A modified version of the standard discrete Radon transform (sDRT) is subsequently applied to each polar representation. An ensemble of classifiers (for detecting curved lines), where each individual (base) classifier is associated with a different polar representation, is therefore constructed. An additional classifier (for detecting straight lines) is obtained through the calculation of the sDRT of the original signature image and is subsequently appended to said ensemble. It is shown that when signatures from different writers are considered, different base classifiers are the most proficient. An average equal error rate (EER) of 9.73% is reported for the majority vote decision of the six base classifiers in question, which represents an effective improvement of 4.37% on the average EER of the most proficient individual classifier.