Online and off-line, automatic signature verification has been an active research topic for decades. Given carefully designed features and curated data, we usually see satisfactory performance, with reductions in equal error rate (EER) to ten percent or even lower. However, when it comes to high-quality forgeries and disguised signatures, which have drawn interest recently, the performance drops significantly.

The authors of this paper attempt to predict the system performance based on several assumptions about the feature normality and the number of samples available. Specifically, the authors address two local features, baseline and loops, and assume (a) the feature properties are normally distributed, (b) the means of original and forged feature values are equal, and (c) the false acceptance rate (FAR) and the false recognition rate (FRR) are equal for each single feature. This allows them to derive a mathematical equation for predicting the lower bound of the average error rate of a signature verification system. Experiments on the public datasets SVC and GPDS-300 show convincing results on normality tests and prediction accuracy. In addition, the prediction on GPDS-300 is more precise than on SVC.

This paper, as it stands, is a nice piece of work. However, I would have liked to see a discussion on how this approach fits in the larger picture of signature verification, especially when forensic handwriting examiners (FHEs) are involved. For example, is it possible to ask FHEs to help calibrate the consistency of features?