非受控场景下单样本人耳识别研究

发布时间：2018-04-21 20:21

本文选题：人耳识别 + 非受控场景　；参考：《北京科技大学》2017年博士论文

【摘要】：人耳识别是最常见的生物特征识别技术之一,具有不受表情影响、不受年龄影响、无需被采集对象配合、可远距离完成等优点。鲁棒的人耳识别系统在诸多方面都有巨大的应用前景,例如门禁管理、出入境管理、法律实施和刑事侦查等方面。经过学术界多年的研究,人耳识别技术已经取得了长足的进步。但是,在学术界将研究重点放在如何提高识别率的同时,一些潜在的根源问题却被忽略了。其中,单样本问题就是亟待解决的问题之一。单样本问题是在现实应用中经常会遇到的,例如抓捕无犯罪前科的嫌疑犯时,公安机关所掌握的可能仅有其一幅证件照片或监控截图。从本质上来看,单样本问题并不仅仅是训练样本数量的问题,而是已知的个体信息不能完全包括待识别样本各种情况下的信息,也就是信息不对称的问题。在非受控场景下,就归结为由姿态、遮挡等情况导致的部分数据问题。针对这样的问题,本文进行了以下三个部分的研究:1)提出一种加权多关键点稀疏表示分类方法(WMKD-SRC:Weighted Multikeypoint Descriptor Sparse Representation-based Classification),用于单样本人耳识别。通过对待识别样本的所有关键点自动施加自适应的权重,进而减小了同一类样本间的类内差别。实验结果表明,该方法能够提高非受控场景下单样本人耳识别的识别率,尤其是在存在遮挡和姿态变化的情况下,能够取得更好的鲁棒性。2)为了增大原型库中各类之间的类间差别,提出一种局部特征权重优化方法(WOLF:Weight Optimization of Local Features)。该方法基于群集智能的经典算法,对原型库中各个样本的局部特征的权重进行优化计算,最终作用于上一部分的识别方法中。实验结果表明,经过权重优化后的方法对姿态变化体现出了更好的鲁棒性。3)为了进一步提高单样本人耳识别的性能,提出一种二维三维数据的决策层融合方法(Hybrid MKD-SRC:Hybrid Multikeypoint Descriptor Sparse Representation-based Classification)和一种二维三维数据的特征层融合方法(TDSIFT:Texture and Depth Scale Invariant Feature Transform)。决策层的融合方法利用由二维纹理图和由三维深度图所获取的信息在同一框架下进行识别,而特征层的融合方法则是提出一种融合二维纹理图和三维深度图的局部特征描述子。实验结果表明,两种方法能够有效地提高单样本人耳识别的识别率,并且与其它方法相比,缩短了计算时间。本文的研究不仅对解决非受控场景下的单样本人耳识别具有重要的研究意义,而且对于类似情况下的其他生物特征识别研究也具有参考和借鉴价值。本文提出的识别方法对于解决现实应用中诸如涉密安保、司法认证、公安机关破获刑事案件等方面具有理论指导意义。
[Abstract]:Ear recognition is one of the most common biometric recognition techniques. It is not affected by facial expression, is not affected by age, does not need to be collected to cooperate, and can be completed from a long distance. Robust ear recognition system has great application prospects in many aspects, such as access control, entry and exit management, law enforcement and criminal investigation. After years of academic research, ear recognition technology has made great progress. However, while the academic community focuses on how to improve the recognition rate, some potential root problems are ignored. Among them, the single sample problem is one of the problems to be solved. The problem of single sample is often encountered in practical application, for example, when arresting a suspect with no criminal record, the public security organ may have only one document photograph or surveillance screenshot. In essence, the problem of single sample is not only the problem of the number of training samples, but also the problem that the known individual information can not include all kinds of information in the case of the sample to be identified, that is, the problem of information asymmetry. In an uncontrolled scenario, it can be attributed to some data problems caused by posture and occlusion. In order to solve this problem, this paper presents a weighted multi-key point sparse representation classification method, WMKD-SRC: weighted Multikeypoint Descriptor Sparse Representation-based classification, which is studied in the following three parts: 1) for single sample human ear recognition. By automatically applying adaptive weights to all the key points of the identification samples, the intra-class differences between the same samples are reduced. The experimental results show that this method can improve the recognition rate of ear recognition of uncontrolled scene samples, especially in the presence of occlusion and posture change. In order to increase the differences between classes in the prototype library, a local feature weight optimization method is proposed. Based on the classical algorithm of cluster intelligence, this method optimizes the weights of local features of each sample in the prototype library, and finally acts on the recognition method of the previous part. The experimental results show that the weight optimization method shows better robustness to attitude change. 3) in order to further improve the performance of single sample human ear recognition, This paper presents a decision level fusion method for 2D 3D data, Hybrid MKD-SRC:Hybrid Multikeypoint Descriptor Sparse Representation-based Classification, and a feature layer fusion method for 2D 3D data, which is TDSIFT: and Depth Scale Invariant Feature transform. The method of decision level fusion uses the information obtained from 2D texture image and 3D depth map to be recognized under the same frame, while the fusion method of feature layer is to propose a local feature descriptor for fusion of 2D texture image and 3D depth map. The experimental results show that the two methods can effectively improve the recognition rate of single sample human ear and shorten the computing time compared with other methods. The research in this paper not only has important significance to solve the problem of single sample ear recognition in uncontrolled scenarios, but also has reference value for other biometric recognition in similar situations. The identification method proposed in this paper is of theoretical significance in solving practical applications such as security, judicial authentication, and the detection of criminal cases by public security organs.
【学位授予单位】：北京科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】