【作者】 于贵平；

【导师】 李德昌；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2004， 硕士

【摘要】 自动指纹识别技术作为生物识别的一个重要组成部分在近年来一直倍受人们的关注。从九十年代中期到现在有很多算法被提出，并出现了一些商业产品，但是因为自动指纹识别技术本身的难度大，加上技术保密，要实现一个实用的指纹自动识别系统非常困难。本文在大量科研实践的基础上，对指纹自动识别技术的各个环节进行了深入研究，提出了一套创新算法，并在此基础上研制开发了一套高效的指纹自动识别系统。指纹识别大致分为指纹图像预处理、特征提取、纹型分类、比对和指纹图像压缩五部分。指纹图像预处理目的是去除指纹图像中的各种噪声，连接脊线的断裂，消除指纹图像形变,恢复清晰完整的指纹脊线结构。首先，根据指纹图像在各个方向上的梯度信息计算局部脊线方向，这里引入噪声图来表示局部脊线的质量，计算脊线方向时，所选取的窗口大小对计算出的方向影响很大，窗口大时可以很好的抑制噪声，但得到的方向精度不足；窗口小时，方向精度好，也容易受噪声的干扰。我们可以取长补短，把各种窗口的优点结合到一起，结合策略是以噪声图为依据，在可忍受的噪声下尽量选取小的窗口计算，这样就达到最高的精确性，在噪声超过预先设定的阈值时，换用更大的窗口计算方向。引入局部脊线频率来刻画了脊线的疏密程度，局部脊线的频率是指在方向场垂直方向上单位长度中脊线数的倒数。在以上计算的基础上，我们用Gabor滤波器来处理指纹图像，Gabor变换就是使用了Gaussian函数作为窗函数的窗口Fourier变换。它具有优异时空特性，我们设计Gabor滤波器使其在频率域的频率与指纹的局部脊线频率一致，使Gabor滤波器在空间域的方向与指纹图像的局部方向一致，这样设计的滤波器大大提高了指纹图像的增强效果。对Gabor滤波后的指纹图像二值化我们采用基于脊线频率的方向二值化，按照脊线的局部方向和频率来设置二值化的计算窗口，使窗口的方向与脊线的方向一致，窗口的大小可以包含两个纹线周期。用窗口内的均值为阈值进行二值化，细化工作相对简单，直接采用A Datta ,S.K Parui提出的并行细化算法对二值图像进行细化。本文提取的特征包括：奇异点、特征点间的脊线数、细节特征点和影子点。奇异点的提取使用基于Poincare index的奇异点提取算法，此算法非常高雅、有效，不过它提出的奇异点对噪声敏感，我们用称之为“向心强度系数”的方法来矫正奇异点。如果图像上某点的向心强度系数大说明此点周围的方向是包含各种方向且按大小依次出现。在方向图上观察该点好像处于涡旋的中心。假奇异点的向心强度系数会明显的小于真正特征<WP=64>点。提取脊线数特征时，先用“Bresenham直线生成算法”在两个特征点之间生成直线，理想的线相交只有一个交点，但在象素级别有很多中可能情况。出现相交时可能有：0，1或多个交点。应用本文提出的“直线与脊线求交算法COLAL（cross of line and line）”，在出现无交点的情况时，进行补点以避免无交点的情况发生，多交点的情况发生在直线斜率和指纹脊线的斜率非常接近的情况下，对此依然应用补点算法。除此之外保存两个点的缓冲，分别存放当前交点和上一个交点，当求得的交点连续或距离很小的情况下，则消去上一个交点。细节特征点提取时，我们引入交叉数Cn(p)表示p点8连通域中脊线与谷线交替变化的次数。引入脊线点数Sn(p) 表示p点8连通域中脊线的条数。脊线上点的交叉数和脊线点数有以下几种情况：1、Cn(p)=1、Sn(p)=1,称为末梢点；2、Cn(p)=2、 Sn(p)=2或3或4,称为连续点；3、Cn(p)=3、Sn(p)=3,称为分支点；据此我们可以提取细节特征点。由于存在毛刺、眼、桥等假特征，我们需要用脊线跟踪技术来去除假细节特征点，滤除办法：1、如果跟踪到末梢点，且跟踪的距离小于一定的阈值，则可以认为跟踪到毛刺，然后把跟踪过的脊线抹去，并把两端的分叉点和末梢点删除。2、如果跟踪到的点中出现重合的分叉点时，则可认为跟踪到的眼，删除眼和眼两端的分叉点。3、当两个分叉点被相互跟踪到时，判断分叉点的另外两个点所在的直线是否和对应的分叉点的相互平行，如果大致平行则可认为跟踪到桥，把跟踪的桥抹去，并删除桥两端的分叉点。我们引入影子点，它与细节特征点伴生，它是从细节特征开始沿脊线跟踪特定距离得到的附加特征点。它的作用是增加细节特征点的局部脊线结构信息。在小公共区比对和现场比对中有很重要的意义。利用脊线跟踪技术可以方便的提取影子点。本文提出基于脊线流向统计的纹型分类算法。脊线的流向是指脊线的旋转和延伸方向，这些方向可以从方向场中提取出来，我们定义几种脊线流向量：下旋流，上旋流，左旋流，右旋流，近左箕的涡旋流，近右箕的涡旋流，次上旋流。根据脊线流向量设定分类规则，就可以分出指纹的基本类别。同一指纹不同次的捺印可能出现平移、旋转、小公共区、非线性形变等情况，加上手指的自然状况差，比对两枚指纹的难度是非常大的。本文介绍一种基于特征点的比对算法，算法设计目标是解决上述所列举的比对中存在的困难。我们在极坐标系中描述特征点集合，这样可以解决平移和旋转带来的影响。指纹图像的非线性形变往往呈放射状，在极坐标中能更好地描述非更多还原

【Abstract】 Recent years, Fingerprint recognition as an important part of Biometric recognition catch people’s attention all along. From 1960s to now, many fingerprint recognition algorithms have been designed, and a number of biometric products have been developed, but fingerprint recognition is a complex pattern recognition problem, further more, some technology is kept secret for commercial interest, to implement a high-powered AFIS (Fingerprint Identification System) is still a very challenge task. In this paper, all the progress made in automatic techniques for fingerprint recognition are been analyzed, a suit of innovated approaches and an effective fingerprint recognition system have been designed.Fingerprint recognition mainly includes five parts: Fingerprint Representation, Feature Extraction, Fingerprint Classification, Fingerprint Matching, and Fingerprint Image Compress.Fingerprint Representation aims at getting rid of the noise in the fingerprint image, connecting the rupture of ridge and eliminating fingerprint image’s distortion. First, we estimate the local ridge orientation and local ridge noise by the local gray grads at different orientation. We use a window to compute local gray grads, but the size of the window will affects the result. Using bigger window can put down the bad effect caused by noise for local ridge orientation, but the local ridge orientation we get is low precision. Otherwise, using smaller window can put down the bad effect caused by noise for local ridge orientation, but the local ridge orientation we get is high precision. We can give a good balance, according to the local ridge noise, if with the noise, the smaller window can be used, we can get a precise orientation. If not, the bigger window is adopted and we get the most precise orientation. We import Local Ridge Frequency describing how dense the ridge is. We use Gabor Filter to enhance fingerprint image. Gabor transform is a Fourier transform which window function is Gaussian function. Gabor transform has good character in time and frequency territory. We make Gabor Filter’s frequency accords with the local ridge frequency and orientation accords with the local ridge orientation. With these setting, Gabor Filter can improve fingerprint image markedly. For turning improved fingerprint image to binary image, we use special filter window that holds more than 2 local ridges and takes the same orientation as the local ridge. We take the parallel thinning method presented by A Datta and S.K Parui to get thin image based on binary image.The characters we detection include Singularity, Minutiae, Ridge Number and Shadow points. The algorithms based on Poincare Index can <WP=67>detection Singularity most decently and effectively, but it is affected by the noise. We use “centrality intensity coefficient” to correct the Singularity. The true Singularity’s centrality intensity coefficient is especially larger than the false one. For Minutiae detecting, first we shall create line between the two characters by the algorithms presented by Bresenham, In practice, the cross number of two curve can be 0 or more than 1. In this paper, the algorithms of COLAL (cross of line and line) are presented. It used can patch the skipped point and eliminate the continuous points. For detecting Minutiae, the cross number Cn(p) and ridge number Sn(p) are imposed, Cn(p) is the change number of ridges and valleys around the point p. Sn(p) is the ridge number around the point p. When Cn(p)=1、Sn(p)=1,p is ridge endings; When Cn(p)=2、 Sn(p)=2或3或4,p is a continuous point；When Cn(p)=3、Sn(p),p is ridge bifurcations. Because the existence of false characters such as burrs, eyes, bridges, we should utilize the techno-ridge-tracing to eliminate false minutiae, according to techno-ridge-tracing do the follow: 1. when tracing the ridge endings and the tracing distance being smaller than a threshold, there would be a burr, wiping off the current point and burr. When the superposition ridge bifurcations occur, there would be an eye between the two the current s更多还原