Srihari Sargur, Huang Chen, Srinivasan Harish
Department of Computer Science and Engineering and Director, Center of Excellence for Document Analysis and Recognition, University at Buffalo, State University of New York, Buffalo, NY 14228, USA.
J Forensic Sci. 2008 Mar;53(2):430-46. doi: 10.1111/j.1556-4029.2008.00682.x.
As handwriting is influenced by physiology, training, and other behavioral factors, a study of the handwriting of twins can shed light on the individuality of handwriting. This paper describes the methodology and results of such a study where handwriting samples of twins were compared by an automatic handwriting verification system. The results complement that of a previous study where a diverse population was used. The present study involves samples of 206 pairs of twins, where each sample consisted of a page of handwriting. The verification task was to determine whether two half-page documents (where the original samples were divided into upper and lower halves) were written by the same individual. For twins there were 1236 verification cases -- including 824 tests where the textual content of writing was different, and 412 tests where it was the same. An additional set of 1648 test cases were obtained from handwriting samples of nontwins (general population). To make the handwriting comparison, the system computed macro features (overall pictorial attributes), micro features (characteristics of individual letters), and style features (characteristics of whole-word shapes and letter pairs). Four testing scenarios were evaluated: twins and nontwins writing the same text and writing different texts. Results of the verification tests show that the handwriting of twins is less discriminable than that of nontwins: an overall error rate of 12.91% for twins and 3.7% for nontwins. Error rates with identical twins were higher than with fraternal twins. Error rates in all cases can be arbitrarily reduced by rejecting (not making a decision on) borderline cases. A level of confidence in the results obtained is given by the fact that system error rates are comparable to that of humans (lower than that of lay persons and higher than that of questioned document examiners [QDEs]).
由于笔迹受到生理、训练和其他行为因素的影响，对双胞胎笔迹的研究可以揭示笔迹的个体性。本文描述了这样一项研究的方法和结果，其中通过自动笔迹验证系统对双胞胎的笔迹样本进行了比较。这些结果补充了之前一项使用不同人群的研究结果。本研究涉及206对双胞胎的样本，每个样本由一页笔迹组成。验证任务是确定两份半页文档（原始样本被分成上下两半）是否由同一个人书写。对于双胞胎，有1236个验证案例——包括824个书写文本内容不同的测试，以及412个书写文本内容相同的测试。另外从非双胞胎（普通人群）的笔迹样本中获得了1648个测试案例。为了进行笔迹比较，该系统计算了宏观特征（整体图像属性）、微观特征（单个字母的特征）和风格特征（整个单词形状和字母对的特征）。评估了四种测试场景：双胞胎和非双胞胎书写相同文本和书写不同文本。验证测试结果表明，双胞胎的笔迹比非双胞胎的笔迹更难区分：双胞胎的总体错误率为12.91%，非双胞胎的为3.7%。同卵双胞胎的错误率高于异卵双胞胎。通过拒绝（不就）临界案例做出决定，可以任意降低所有情况下的错误率。系统错误率与人类相当（低于外行人员且高于文件检验专家[QDEs]）这一事实表明了对所获结果的置信度。
