Persian offline signature verification based on curvature and gradient histograms

The distinctive characteristic of Persian signature implies its own specific feature extraction approaches for verification systems. In this paper, inspired by the cursive nature of Persian signature, and motivated by the opinion of Persian forensic handwritten experts that put emphasis on this factor, we propose a framework for Persian offline signature verification that uses histogram of curvature (HOC) and histogram of oriented gradient (HOG) as features. Calculating curvature for offline signature due to lack of temporal information is not a straightforward task. To this end, we use a discrete curvature estimator that works by normalized gradient and its Jacobian matrix. We investigate the effect of preserving signatures' scale, center of gravity, grid size, grid overlapping, number of bins, and smoothing factor on the accuracy. Support vector machine (SVM) is used to train writer-dependent classifiers on genuine signatures and random forgeries as training samples. Results show competitive performance between HOC and HOG descriptors on UTSig dataset which consists of 8280 samples from 115 classes. Using combination of HOC and HOG descriptors, we achieve promising equal error rate for finding genuine signatures versus skilled forgeries, while false random forgery acceptance is close to zero.