基于全体数据分布规律的指纹证据似然比评价方法研究

Research on the Evaluation Method for Likelihood Ratio of Fingerprint Evidence Based on the Distribution Law of Entire Data

为解决千万人级数据库中指纹自动识别系统比对得分全体数据难以获取的问题,构建科学有效的指纹证据似然比模型,需要对全体数据的分布规律进行探索。以指纹查询比对得分为数据基础,通过R语言采用矩估计和极大似然估计等数理统计方法,对Beta、Norm、Weibull和Gamma 4种参数估计分布函数进行数据拟合和误差计算,以确定最佳的参数估计方法。根据最佳拟合参数方法模拟全体数据,构建似然比模型并计算同源和异源似然比对数值以直方图评估其性能。实验结果表明,Beta分布在单枚和多枚指纹的不同细节特征数量分布规律中误差值最小,拟合效果最好。分段和整体考虑二者细节特征数量所构建的模型相似,随特征数量从5至16个的增加,同源和异源得分数据的概率密度函数曲线分离程度变大,模型性能变强,尤其当细节特征数量超过12个后,同源和异源曲线几乎完全分离,直至16个细节特征时都表现出极好的区分识别能力。针对单枚指纹所分的三段模型与指纹鉴定实际工作相符,但考虑所有特征数量时,会因综合数据较多而使模型性能下降。Beta分布可较好地展现异源条件下全体指纹数据的分布规律,以此构建的似然比模型的区分识别能力较强,具有很好的实用前景,有助于推动指纹证据评价从经验走向科学。

The distribution law of the entire data is necessary to be explored to construct a scientific and effective likelihood ratio model of fingerprint evidence for solving the problem of difficult access to the en⁃tire data of the comparison score of automatic fingerprint identification system in the ten⁃million⁃people database.Taking the fingerprint comparison scores as the data base,the data fitting and error calculation of four parameter estimation of distribution functions,including Beta,Norm,Weibull and Gamma,are carried out through the mathematical statistical method in R language using moment estimation and maxi⁃mum likelihood estimation,so as to determine the optimal method of parameter estimation.The entire da⁃ta is simulated by the best⁃fit parameter method,the likelihood ratio model is constructed and logarithmic values of same⁃source and different⁃source likelihood ratios are calculated to assess its performance in his⁃tograms.Experimental results show that Beta distribution has the smallest error value and the best fitting effect in the distribution law of the number of different minutiae for both single and multiple fingerprints.The models constructed by sectional and overall analysis of the number of minutiae are similar.As the number of minutiae increases from 5 to 16,the separation degree of probability density function curves for same⁃source and different⁃source score data increases,and the performance of the models becomes bet⁃ter.Especially when the number of minutiae exceeds 12,the same⁃source and different⁃source curves are almost completely separated,showing excellent differentiation and recognition ability up to 16 minutiae.The three⁃segment model divided for single fingerprints is consistent with practical work in fingerprint i⁃dentification.However,when considering the number of all minutiae,the performance of the model would be degraded due to the large amount of integrated data.Beta distribution can better present the dis⁃tribution law of the entire fingerprint data under different⁃source conditions,and the likelihood ratio mod⁃el constructed by this way has strong identification ability,which has a great practical prospect,and helps to promote the evaluation of fingerprint evidence from experience to science.