【作者】 高修峰；

【导师】 张培仁；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 自动人脸识别在公共安全、智能监控、数字身份认证、电子商务、多媒体和数字娱乐等领域具有巨大的应用价值,同时,人脸识别的研究涉及多个学科,具有重要的理论研究价值,受到各国政府、科研单位以及军事、安全、情报部门的广泛关注和高度重视。经过几十年的研究,人脸识别取得了长足的发展与进步,目前在控制和配合条件下,人脸图像识别可以取得比较高的识别率,在一定场合下,已经可以应用到人们的生活中;但是在非控制条件和非配合条件下的人脸图像识别仍然是一个极具挑战性的课题。由于人脸识别的广泛应用,ISO/IEC成立工作组制定相关的标准,其中之一就是评估对系统性能造成影响的各种因素。2006年,中科院自动化所生物识别与安全技术研究中心代表中国向ISO/IEC标准委员会提交关于人脸图像质量的标准草案,该标准与2007年2月确立初稿,经修改,讨论与2007年5月8号通过。本文作者作为该标准的重要参与者之一,参与了该标准的起草修改过程,本文对该标准的做了详细介绍并对提出了一种新的基于人脸对称性的非对称光照和姿态评估方法。在人脸图像质量标准提交之前,指纹图像标准已经被公布,由于指纹是一种接触式的生物特征,采集图像的条件大都可以控制,因此指纹图像质量评估更注重特征的有效性,因此标准的制定相对比较直观;但对于人脸识别,由于人脸属于非接触性生物识别,这也造成了影响人脸图像质量的因素的多样性,如何确立标准对人脸图像的各个因素进行评估是相对很困难的。此外,对于一幅人脸图像,如何给出一个总体的评价,尚无相关的理论和算法,这也对标准的确立带来很多麻烦。针对这一问题,本文提出人脸图像质量评估的标准方法框架。本文首先定义质量评估方面的一些概念,然后对影响图像质量的各种因素做了总结和分类。针对各个影响图像质量的因素,本文给出了相关的评估方法。对于图像的整体特性,如全局光照,对比度等,前人已做了相关的评估研究,本文主要总结了一些评价这些因素的依据,从而为以后的算法确立一定的参考和依据,这些依据也得到的国际专家的认可,已被收录到ISO/IEC国际标准中。本文对这些因素对识别系统的影响作了定量的实验。鉴于影响人脸图像质量因素的多样性,且相关的评估算法均为对单一因素做评估的现状,最后,本文采用了两种分数:各因素的单一质量分数和图像的整体分数来对图像质量进行量化的表示。前人的研究表明,光照和姿态是影响人脸识别系统的两个最为重要的因素,虽然针对这两个问题,人们提出了很多算法来削弱光照和姿态对识别系统的影响,但始终没有很大的突破,无法从根本上解决这一问题,因此,ISO/IEC在制定人脸图像标准时也需要为这两个问题作出全面客观评估。针对这一问题,本文提出采用人脸的对称性这一人脸的统计特征来评估非对称光照和不正确姿态,为了更好的描述人脸的对称性,本文借鉴于LBP子窗口的思想,采用局部子窗口来描述人脸左右半边的对称性。文中给出了对称性的评价公式,然后根据该公式对光照和姿态进行了初步的评估,该方法已被收录到ISO相关的标准中。ISO相关标准指出,生物图像质量评估的最终目的是最大化图像质量与系统性能即匹配引擎输出之间的联系,由于目前并没有一个标准的带有图像质量的数据库。所以大多的图像质量评估算法的标准质量均有人为来划定,这并不符合图像质量评估的目的。本文针对这一问题,首先通过实验证明,采用人为判定的图像质量与系统输出的匹配分数并没有线性的关系,即图像质量高并不意味着图像的正确匹配分数高(所谓正确匹配分数就是同一个人的两幅图像匹配产生的分数),而图像质量低也不意味着图像的正确匹配分数低。作为改进,本文提出采用回归的方法在图像与匹配分数之间建立联系,木文研究了两种回归方法,一种是传统的线性回归,一种是基于boosting的非线性回归方法。木文采用训练测试的机制,首先通过训练在图像和匹配分数之间建立模型,然后通过测试验证该模型的正确性,实验证明,本文提出的方法可以有效的预测人脸图像的匹配分数,这也对以后改进系统性能提供了一个很有意义的基础。更多还原

【Abstract】 Automatic face recognition has great potential applications in public security, intelligent surveillance, digital personal identify, electronic commerce, multimedia, digital entertainment, etc. and has great theory value in many subjects, so face recognition has attracted much research attention from the research institutes, governments, military and security departments. Over the past 30 years, great progress and developments have been made in face recognition. Now under the controlled and cooperative conditions, face recognition systems perform very well, but under uncontrolled and uncooperative conditions, especially when the illumination in face images and facial poses are variant. For the wide use of face recognition, ISO/IEC established a group to draft standards about face image. One of them is to evaluate the aspects which influent system performance.In 2006, CBSR(Center for Biometrics and Security Research) summit a standard draft on face image quality to ISO/IEC on behalf of China. In February 2007, the working draft was published. After some discussion and modification, the standard was passed in May 8. 2007. The author of this article, which is partner of this standard, will give a detailed introduction on this standard in this article. Also this paper proposes some algorithms to evaluate non-frontal lighting and improper facial pose.Before the standard was summit, some standards about fingerprint quality were established. Fingerprint is a touchable biometric feature, the conditions of capturing fingerprint image are under control. So fingerprint image quality evaluation is mostly about the effect of features. But for face recognition, the aspects that influent image quality are diversify because face recognition is a un-touchable biometric feature. So how to establish a standard for face image is difficult. On the other hand, for a face image, how to give an overall evaluation, there are no research on this problem. For this problem, this paper proposes a standard framework to evaluation the aspects that affect face image quality. First, some definition about quality evaluation are defined. Then we classify these aspects into some class. For most of the aspects, we give the evaluation methods in this paper. This gives a basement for future work on face image quality evaluation. Until now, most evaluation algorithm is to evaluation one aspects, so we propose an aspects-score and over-score method to give face image an overall evaluation.Based on the past research, non-frontal lighting and improper facial pose are the most important aspects that affect the system performance. For these two aspects, researchers proposes many algorithms to weaken the influence of them, but no big progress was made. These two problem still have not been solved. So the standard need to give evaluation for these problems. For this reason, this paper propose a method which use symmetry, a statistical feature of face to evaluate non-frontal lighting and improper facial pose. To describe symmetry, this paper use local windows to evaluate the symmetry of face image. This idea is came from LBP. After evaluation of symmetry, we give the methods to assess non-frontal lighting and improper facial pose.From the ISO stands, image quality must be directly connected with matching performance. But until now, there are no database with standard image quality. So most image quality methods used the quality marked by hand, this is not proper. For the problem of image quality can’t been directly connected with matching performance, this paper uses regression method to build models between face image and genuine matching score. For the face images with non-frontal lighting and pose variation, first we evaluate the input image’s symmetry, then build models between face image symmetry and matching score. Two regression methods are discussed, the first one, which selects some most effective features using Adaboost, build a linear model between these features and matching score. The second one is a non-linear method based on boosting. The experiment result show that these two methods can predict matching score very well.更多还原