血管类图像分割与识别方法研究

发布时间：2018-05-30 09:43

本文选题：特征提取 + 手指静脉识别　；参考：《山东大学》2016年博士论文

【摘要】：生物特征识别相较于传统的身份验证具有安全、方便等优势,其研究价值和良好的应用价值得到了学术界和企业界的认可。血管作为藏匿在人体皮肤下的模态,具有活体识别、不易被盗用和复制等特点,是更加安全的一类生物特征。血管类的模态主要包括以手指静脉为代表的手部血管和眼部视网膜血管等,本文针对手指静脉识别和视网膜血管分割、识别中存在的一系列问题进行了研究。手指静脉识别是一种比较新颖的、被业内认为是最有前景的身份认证技术之一。相较于其他的生物特征识别方法,手指静脉识别具有安全性高、方便和用户友好等优势,得到了越来越多同行的关注。目前对手指静脉识别的研究取得了较大的进展,但仍有问题需要解决。例如,手指静脉图像存在的图像质量和形变问题,会对基于局部特征手指静脉识别方法的效果产生较大的影响。改进现有的特征,以及设计鲁棒的特征提取方法是解决这些问题的重要方向。视网膜眼底图像主要包括眼底血管和感光细胞等结构。视网膜眼底血管是人体血液循环系统唯一可以无创伤观察的较深层微血管。视网膜眼底图像可以用于身份验证,此外,视网膜眼底血管的特征对诊断系统疾病和系统性眼病也有重要意义。无论进行视网膜识别还是对视网膜血管特征进行分析,一般都需要对血管进行分割。在视网膜血管分割方面存在大量工作,但在效率和准确率上很难达到一个平衡,如何高效、准确的进行血管分割仍是一个挑战。在利用视网膜进行身份验证时,不准确的血管分割易引入错误,如何在避免血管分割的情况下进行视网膜识别也是一个重要研究课题。本文针对手指静脉识别中局部特征的表达能力差和对形变敏感等问题,以及如何高效、准确的进行视网膜血管分割、如何在避免血管分割的情况下进行视网膜识别进行了深入的分析和探讨,主要工作和贡献有：1、基于局部特征的识别是手指静脉识别中一类比较重要的方法。比较常用的局部特征主要有局部二值模式(LBP)、局部导数模式(LDP)及其变种等。现存的局部特征主要考虑到了像素邻域内的梯度方向,而忽略了梯度的大小和梯度之间的关系等,所以特征的表达能力有限。基于此分析,本文设计了一种新的局部特征提取方法,称为局部方向编码(LDC),该方法不仅考虑了局部梯度变化的大小,且进一步考虑了多个方向的梯度变化关系。在包含136个手指的4,080幅图像上的识别效果表明了所设计特征的区分能力。本文方法比目前最优的基于局部特征的LLBP方法的等错误下降了50%。2、目前的手指静脉识别技术在处理形变问题时,往往将其视作是一种影响匹配的噪声信息,并将工作的重点放在如何对形变信息进行恢复或者怎样降低形变的影响上,而忽略了形变信息本身的规律性。经分析,在进行同源匹配时,虽然两幅图像间存在形变,但是由于像素位置关系的约束,相邻像素的位移和方向是相似的；而在异源匹配时,相邻位置的像素由于特征的差异,位移和方向就没有这种规律。基于此,本文提出了基于形变信息的手指静脉识别方法,利用匹配中形变矩阵的一致性来衡量两幅图像的相似度。形变矩阵由基于像素级特征的优化匹配算法产生。在公开手指静脉库PolyU和SDU-MLA上六折交叉验证的等错误率分别为0.0010和0.0049,表明了所设计特征的区分能力以及对识别的有效性。3、现有的视网膜分割算法主要分为监督方法和非监督方法两类。监督方法的分割效果较好,但是需要提取像素级的特征,所以效率较低；而非监督方法由于提取的特征较简单,得到的效果往往不理想。此外,这两类方法对边缘像素的区分性都达不到理想的效果。本文设计了一种基于滤波器的非监督视网膜血管分割方法,主要思想是增强血管和背景之间的差别,简化分割过程。首先,设计两种不同的滤波器,一种增强图像的细节,另一种对整幅图像的光照进行归一化,然后对两种滤波的结果进行加权得到最终的效果。这样得到的预处理图像只需要一个阈值就可以将血管分割出来。经过简单的微调和后处理之后,在公开数据库DRIVE和STARE库上的实验结果超过了目前的非监督方法,且优于大多数监督方法,并有较高的效率。4、SIFT描述子以其强大的区分能力著称,且在物体识别和检测上得到了广泛应用。经验证,SIFT描述子在视网膜识别的直接应用却得不到理想的效果,在VARIA数据库上的等错误率为0.0436。本文认为找到效果不理想的原因,并提高基于SIFT的视网膜识别的准确率是一个值得研究的问题。经分析发现,该类图像往往存在光照不均匀、模糊和局部对比度小等问题,这样的图像在进行SIFT特征提取时,继续进行高斯子空间变换,会损失大部分图像信息,提取的SIFT描述子也会不稳定。基于上述分析,本文对视网膜图像进行了去除偏场的增强操作,并进一步利用迭代的各向异性扩散平滑算法对图像进行去噪。在经过预处理的图像上,基于SIFT的识别准确率有明显提高且有较强的鲁棒性。
[Abstract]:Compared with traditional identification, biometric identification has the advantages of security, convenience and other advantages. Its research value and good application value have been recognized by the academia and the business community. As a mode hidden under human skin, blood vessels have the characteristics of living identification, not easy to be embezzled and copied, and are a safer biological feature. The mode of the class consists mainly of the hand vessels and retinal vessels represented by the finger vein. In this paper, a series of problems in the recognition of finger veins and retinal vascular segmentation are studied. Finger vein recognition is one of the most promising and most promising identification techniques in the industry. Compared with other methods of biometric identification, finger vein recognition has the advantages of high security, convenient and user friendly. More and more attention has been paid to the research of finger vein recognition. However, some problems still need to be solved. For example, the image quality and deformation of the finger vein image can be solved. It has a great influence on the effect of finger vein recognition based on local characteristics. Improving the existing features and designing a robust feature extraction method are important directions to solve these problems. Retina fundus images mainly include fundus blood vessels and photosensitive cells. Retina fundus blood vessels are the only blood circulation system of the human body The retina fundus images can be used for identification. In addition, the characteristics of retinal fundus vessels are also important for diagnosis of systemic diseases and systemic ophthalmopathy. There is a lot of work in the division of omentum vessels, but it is difficult to achieve a balance in efficiency and accuracy. How to effectively and accurately segment the blood vessels is still a challenge. When using the retina for identification, the inaccurate segmentation of blood vessels is easy to introduce errors, and how to avoid retinal recognition under the condition of avoiding blood vessel segmentation This paper is an important research topic. In this paper, in view of the poor expression of the local characteristics of the finger vein recognition and the sensitivity to deformation, and how to effectively and accurately segment the retinal vessels, how to do retinal recognition in the condition of avoiding blood vessel segmentation is deeply analyzed and discussed. The main work and contribution are as follows: 1, Local feature recognition is one of the most important methods in finger vein recognition. The commonly used local features mainly include local two value mode (LBP), local derivative mode (LDP) and its variant. The existing local features mainly take into account the gradient square in the neighborhood of pixels, but ignore the relationship between the gradient and the gradient. Based on this analysis, this paper designs a new local feature extraction method called local direction coding (LDC). This method not only takes into account the size of local gradient change, but also considers the gradient change of multiple directions. The recognition effect on 4080 images containing 136 fingers. It shows the distinguishing ability of the designed features. This method is less than the 50%.2 of the optimal LLBP method based on the local feature. The current finger vein recognition technology often treats it as a kind of noise information that affects the matching, and puts the focus on how to restore the deformation information. And how to reduce the influence of the deformation, but ignore the regularity of the deformation information itself. After analysis, although the two images are deformed during the homologous match, the displacement and direction of the adjacent pixels are similar because of the pixel position relation, while the pixels in the adjacent position are due to the difference of the features in the source matching. In this paper, a finger vein recognition method based on deformation information is proposed, and the similarity of the two images is measured by the consistency of the deformation matrix in the matching. The deformation matrix is produced by the optimization matching algorithm based on the pixel level features. Forty percent off on the PolyU and SDU-MLA of the finger vein library. The error rates of cross validation are 0.0010 and 0.0049 respectively, indicating the distinguishing ability of the designed features and the validity of the recognition. The existing retinal segmentation algorithms are divided into two categories: supervision method and unsupervised method. The segmentation effect of the supervised method is better, but the feature of pixel level needs to be extracted, so the efficiency is lower than that of the.3. In addition, the two methods are not ideal for the distinction between the edge pixels. In this paper, a non supervised retinal blood vessel segmentation method based on filter is designed. The main idea is to enhance the difference between the blood vessels and the background and simplify the segmentation process. First, two different filters are designed, one is to enhance the details of the image, the other is the normalization of the illumination of the whole image, then the final effect is weighted for the results of the two filtering. The obtained preprocessed image can be divided out of the blood tube with only one threshold. After that, the experimental results on the public database DRIVE and STARE library exceed the current unsupervised methods, and are superior to most supervision methods, and have high efficiency.4. The SIFT descriptor is famous for its powerful distinguishing ability, and has been widely used in object recognition and detection. It is verified that the SIFT descriptor is directly responsible for the retina recognition. It is not ideal for use, and the error rate in the VARIA database is 0.0436.. This paper thinks that it is a problem to find out the reason of unsatisfactory effect and improve the accuracy rate of retina recognition based on SIFT. When SIFT feature extraction is carried out, the Gauss subspace transformation is continued, most of the image information will be lost, and the extracted SIFT descriptor will be unstable. Based on the above analysis, the enhancement operation of removing partial field is carried out on the retina image, and the image is de-noised by the iterative anisotropic diffusion smoothing algorithm. On the preprocessed images, the recognition accuracy based on SIFT has been significantly improved and has strong robustness.
【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】