指纹特征的统计模型及其在潜指纹和重叠指纹处理中的应用研究

Statistical modeling of fingerprint features is a fundamental problem in fingerprint individuality analysis and many other fingerprint image analysis applications, such as orientation field extraction and synthetic fingerprint generation. However, existing fingerprint feature statistical models are not invariant to fingerprint distortion, which seriously degrades the robustness and accuracy of these models because fingerprint distortion is commonly seen in fingerprint images. In this research, we propose to integrate the fingerprint distortion parameters into the models of fingerprint features, and design an effective optimization approach to simultaneously estimate model and distortion parameters. We further investigate the structures among different features, e.g. distances and direction differences between neighboring minutiae, and model their distributions in fingerprints. The resulting feature models, unlike existing fingerprint feature models, take into consideration higher order statistics of fingerprint features, and are thus more accurate in representing fingerprints. Finally, we attempt to apply the obtained fingerprint feature models to solving some emerging challenging problems in automated fingerprint recognition, i.e. latent and overlapping fingerprint processing. More specifically, latent fingerprint segmentation is fulfilled via fitting fingerprint feature models, latent fingerprint feature extraction and matching is improved based on the fingerprint feature model fitting accuracy and by using fingerprint feature models to predict unknown features in missing parts of latent fingerprints, and overlapping fingerprint separation is implemented by reconstructing the orientation fields of overlapping fingerprints according to fingerprint feature models. The proposed models and algorithms are validated and evaluated on a number of fingerprint databases. The achievements of this research, including the novel robust and accurate fingerprint feature models and the novel effective and efficient latent and overlapping fingerprint processing algorithms, are expected to be useful for the understanding of fingerprint evidence in forensics as well as the design and implementation of effective automated fingerprint recognition technology.

自动指纹识别技术在过去几十年中得到了飞速发展，而且由于指纹唯一性分析的需要，不少指纹特征模型被提了出来。但是，现有的指纹特征统计模型对指纹形变不具有不变性，考虑的指纹特征也比较单一，忽略了指纹特征之间的结构关系，因而这些模型的鲁棒性和精确性都较差。另一方面，当前的自动潜指纹识别率仍然远低于指纹专家手工识别的准确率，其主要问题在于潜指纹质量差、图像背景复杂、指纹形变大、且所含有的指纹特征量少。针对这些问题，本项目将指纹形变引入指纹特征模型，利用聚类分析和优化算法同时估算模型系数和形变参数，分析指纹特征之间的结构关系并建立其分布模型，从而提高指纹特征模型的鲁棒性和精确性。此外，本项目将潜指纹和重叠指纹处理问题转化为模型的拟合和泛化问题，并基于所建立的指纹特征模型设计有效的指纹处理和识别算法，以提高特征提取的可靠性、增加有用的特征信息、改进识别率。本课题研究具有学术创新性和重要的应用价值。

自动指纹识别技术经过过去几十年的飞速发展，目前已经在包括刑侦和民用在内的诸多领域中得到了广泛应用。然而，人们对于指纹特征随机性的理解仍然有限，这制约了人们进一步改进指纹识别技术，同时自动指纹识别技术在实际应用中依然面临各种挑战。在本项目的资助下，课题组成员对指纹特征的统计模型进行了深入研究，并利用这些统计模型及其相关方法来解决自动指纹识别技术面临的挑战，取得了一系列重要成果。.本项目取得的成果具体包括：1）我们分析了方向场对指纹细节点特征分布的影响，并据此提出了一种新的细节点统计模型，有效提高了模型对细节点分布的刻画精度；2）指纹奇异点是更基础的一种指纹特征，决定了指纹方向场的基本形态，我们以奇异点为基准建立参考坐标系，将奇异点的影响进一步引入细节点模型，从而更全面更准确地描述了细节点的分布规律；3）指纹形变是影响指纹采集和识别的重要因素之一，因此，我们利用主成分分析和高斯混合模型对三维指纹形状进行了统计建模，并设计算法合成出逼真的三维指纹数据，为评估指纹形变对指纹采集和识别效果的影响提供了基础；4）准确地提取指纹方向场是指纹识别中的关键步骤，为了更好地去除指纹方向场中的噪音，我们引入了字典学习方法来建模指纹方向场，提出了一种有效的指纹方向场提取方法，提高了潜指纹识别率；5）由指纹细节点重构指纹图像在指纹合成和指纹识别系统安全性方面具有十分重要的价值，我们分析了现有指纹图像重构方法生成假细节点的原因，进而通过处理由于奇异点造成的指纹方向场中的不连续性有效降低了假细节点的数目，得到更加精确的重构指纹图像；6）指纹活体检测是自动指纹识别实用系统的重要组成部分，我们提出了基于倒谱分析的指纹图像噪音模型，有效提高了指纹活体检测的准确率，此外，利用深度学习技术，我们还提出了一种基于深度卷积神经网络的指纹活体检测方法，不但具有更高的活体检测准确率，而且速度快，能够满足实时性的要求。.在本项目基金资助下，项目组共发表论文8篇，申请专利1项，其中包括SCI期刊Electronics Letters上论文1篇、SCI期刊Information Science上论文1篇和领域内顶级会议International Conference on Biometrics上论文2篇。培养博士研究生3人，硕士研究生7人。项目按照预定计划顺利完成，经费执行情况也较为合理。