Automatic Quantitative Letter-level Extraction of Features Used by Document Examiners

Abstract

In this paper we examine tile automatic extraction of visual or structural features as used by document examiners in the comparison of handwriting samples. We have extracted between 7 and 14 different features from four letters ("y", "d", "f" and "t"). This analysis was earned out on a total of 3077 letters from 30 different writers. On average these features are extracted with about 88% accuracy and can be used to assess the similarity of different writing samples.