手指静脉图像检索方法研究

发布时间：2018-07-29 07:37

【摘要】：随着生物特征技术的快速的发展,类似于虹膜、指纹、人脸等生物特征因为其本身的独特性和可识别性而已经被广泛的应用在身份识别的领域中。在众多的生物特征识别中,手指静脉特征比起上述提到的生物特征有一些独特的优势,主要集中在以下四点:首先是不必担心因为手指触碰到某些地方面导致手指静脉特征被复制和盗取,因为手指静脉特征存在于皮肤的下层;第二点是静脉特征只能从活体中取得,因此静脉特征具有很高的安全性;第三点是获取手指静脉的装置是非常小巧和便携的;最后是对于大多数人来说,每个人最多能获取到10个手指静脉特征,并且每个手指静脉特征都是不同的,因此手指静脉特征对身份验证来说具有很强的可选择性和灵活性。正因为手指静脉特征具有以上这些优势,手指静脉识别技术吸引了非常多的学者的研究热情并且变成一项越来越有前途的技术。很多相关的科研工作已经汇报了手指静脉识别领域的先进成果,但是在大规模人群的应用场景下,手指静脉图像识别工作的鲁棒性依然存在很大的挑战。在这样的情况下,检索方法对于提高识别效率就变得非常有用。鉴于这样的目的,本文第二章节提供了一种基于分层词汇树二值码搜索路径的手指静脉图像检索方法。具体来说,我们首先将局部的静脉基元通过分层的k-means方法来构造一棵词汇树。每个图像块都在搜索词汇树中最近的叶子节点的过程中的二值码路径表示出来,并且该二值码的每一位是根据搜索路径中是否通过或者跳过某个通讯节点而被标记为1或0。任意两张图片的相似度是由两个路径之间包含相同位数的总数来决定的。对一幅查询图像来说,注册图像中根据得分排名,前t个图像将被选为候选者以此来缩小搜索空间。通过在五个手指静脉库上进行验证,试验结果表明本文提出的方法能有效提高检索的精度和效率。随着不断增长的数据量和使用需求,生物特征的数据库也增长成更大的规模,因此手指静脉图像检索也变成快速验证身份的有效方法。然而当前存在的大多数检索方法都是基于非重叠的矩形图像块中提取的固定范围的特征,因此特征表示能力和静脉特征的局部一致性都被忽视。并且这种弱化的编码(根据预先定义的阈值进行二值化)也会限制检索的性能。为了解决这个问题,本文第三章节提出了一种新颖的手指静脉图像检索框架,该框架基于多尺度超像素特征的相似度保留编码方式。在提出的框架中,一个超像素块中的局部一致的像素用来表示一个特征,并且根据超像素的分类,处于不同类别的超像素特征维数长度是不同的,超像素的分类是根据最低维数的特征的方差来决定的。此外,这种基于特征的压缩和循环的特征编码能够将量化损失降到最低并且保持了多尺度特征和二值码直接的相似性。在六个公开的手指静脉数据库上的实验结果表明,本文提出的基于多尺度超像素特征编码的手指静脉图像检索方法比当前最先进的成果效果都好。
[Abstract]:With the rapid development of biometric technology, biometric features like iris, fingerprint and face have been widely used in the field of identity recognition because of their own uniqueness and identifiability. In many biometric identification, the finger vein features have some unique advantages compared with those mentioned above. Focus on the following four points: first, you do not have to worry about the finger vein characteristics being copied and stolen because the finger touches some local surfaces, because the finger veins are characteristic of the lower layer of the skin; the second point is that the venous features can only be obtained from the living body, so the venous features are highly safe; the third point is to obtain the finger veins. The device is very small and portable; in the end, for most people, each person can obtain up to 10 finger veins, and each finger vein features different features, so the finger vein features have a strong selectivity and flexibility for identification. The finger vein recognition technology has attracted many scholars' enthusiasm and became an increasingly promising technology. Many related research work has reported the advanced results in the field of finger vein recognition, but the robustness of finger vein image recognition still exists in the application scene of large population. In this case, in this case, the second chapter provides a finger vein image retrieval method based on the hierarchical vocabulary tree two value code search path. In particular, we first pass the local venous primitives through the layered K-means square. A word tree is constructed by law. Each image block is represented by the two value code path in the search of the nearest leaf node in the vocabulary tree, and each bit of the two value code is based on or skipping a communication node in the search path and the similarity between the 1 or the 0. meaning two pictures is made up of two paths. For a query image, the first t image will be selected as a candidate to reduce the search space for a query image. The results show that the proposed method can effectively improve the accuracy and efficiency of the retrieval by using the five finger vein library. The growing number of data and the use of requirements, biometric databases have also increased to a larger scale. Therefore, finger vein image retrieval has become an effective method for fast identification. However, most of the existing retrieval methods are based on the characteristics of the fixed range extracted from non overlapping rectangular image blocks, so the feature representation is expressed. The local consistency of capacity and venous features is ignored. And this weakening encoding (two values based on predefined thresholds) also limits the performance of the retrieval. In order to solve this problem, the third chapter proposes a novel framework for the finger vein image retrieval based on the similarity of multi-scale super pixel features. In the proposed framework, in the proposed framework, a locally consistent pixel in a super pixel block is used to represent a feature, and according to the classification of the super pixel, the length of the feature dimension of the super pixel is different in different categories. The classification of the super pixel is determined by the variance of the characteristics of the lowest dimension. In addition, this kind of special pixel is based on the characteristic of the minimum dimension. Characteristic compression and cyclic feature coding can minimize the quantization loss and keep the similarity between the multi-scale feature and the two value code. The experimental results on the six open finger vein database show that the hand finger vein image retrieval method based on multi-scale super pixel feature coding is more advanced than the present one. The results are all good.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】