文献翻译--指纹识别系统

1、Fingerprint Recognition SystemAbstractFingerprint recognition system consist of image preprocessing, features extraction and features matching that runs effectively and accurately on personal computer. The image preprocessing includes noise removal, histogram equalization, global thresholding and ri

2、dgeline thinning which are necessary for the features extraction. Extracted features are then stored in a file for fingerprint matching. Matching algorithm presented here is a simple, fast and accurate. Experimental results for matching are accurate, reliable and fast for implementation using a pers

3、onal computer and fingerprint reader. The proposed fingerprint algorithm can provide an effective way of automated identification and can be extended to other security or identification applications. Further the algorithm can be implanted on a FPGA platform for a real time personal automated identif

4、ication system. Keywords:biometric recognition;histogram equalization;ridge thinning;ridge ending;ridge bifurcation.1. INTRODUCTIONFingerprint recognition systems are termed under the umbrella of biometrics. Biometric recognition refers to the distinctive physiological (e.g. fingerprint, face, iris,

5、 retina) and behavioral (e.g. signature, gait) characteristics, called biometric identifiers or simply biometrics, for automatically recognizing individuals. In 1893, it was discovered that no two individuals have same fingerprints. After this discovery fingerprints were used in criminal identificat

6、ion and till now fingerprints are extensively used in various identification applications in various fields of life. Fingerprints are graphical flow-like ridges present on human fingers. They are fully formed at about seventh month of fetus development and fingerprint configuration do not change thr

7、oughout the life except due to accidents such as bruises or cut on fingertips.Because of immutability and uniqueness, the use of fingerprints for identification has always been of great interest to pattern recognition researchers and law enforcement agencies. Conventionally, fingerprint recognition

8、has been conducted via either statistical or syntactic approaches. In statistical approach a fingerprintfeatsures are extracted and stored in ann-dimensional feature vector and decision making process is determined by some similarity measures. In syntactic approach, a pattern is represented as a str

9、ing, tree 1, or graph 2 of fingerprint features or pattern primitives and their relations. The decision making process is then simply a syntax analysis or parsing process.This paper suggests the statistical approach. Experimental results prove the effectiveness of this method on a computer platform,

10、 hence making it suitable for security applications with a relatively small database. The preprocessing of fingerprints is carried out using modified basic filtering methods which are substantially good enough for the purpose of our applications with reasonable computational time. Block diagram for

11、the complete process is shown in Figure.1.Imapt CaflurcStOffidFcatiircs2. IMAGE PREPROCESSINGFor the proper and true extraction of minutiae, image quality is improved and image preprocessing is necessary for the features extraction because we cannot extract the required points from the original imag

12、e. First of all, any sort of noise present in the image is removed. Order statistics filters are used to remove the type of noise which occurs normally at image acquisition. Afterwards the following image preprocessing techniques are applied to enhance the fingerprint images for matching.2.1 Histogr

13、am EqualizationThis method is used where the unwanted part of the image is made lighter in intensity so as to emphasize the desired the desired part. Figure 2(a) shows the original image and Figure 2(b) histogram equalization in which the discontinuities in the small areas are removed. For the histo

14、gram equalization, let the input and the output level for an arbitrary pixel be i and l, respectively. Then the accumulation of histogram from 0 to i ( 0 & i & 255,0 < k < 255) is given byC(/)= Z H(k)(1)Figure 2(a) Original Ini agewhere H(k) is the number of pixel with gray level k, i.

15、e. histogram of an area, and C(i) is also known as cumulative frequency.Figure 2(b| Histogram Equalized2.2 Dynamic ThresholdingBasic purpose of thresholding is to extract the required object form the background. Thresholding is simply the mapping of all data points having gray level more that averag

16、e graylevel. The results of thresholding are shown in Figure 3 .Figure 3 Result of Dynamic Threshold!ng2.3 Ridgeline ThinningBefore the features can be extracted, the fingerprints have to be thinned or skeletonised so that all ridges are one pixel thick. When a pixel is decided as a boundary pixel,

17、it is deleted directly form the image 3-5 or flagged and not deleted until the entire image been scanned 6-7. There are deficiencies in both cases. In the former, deletion of each boundary pixel will change the object in the image and hence affect the object symmetrically. To overcome this problem,

18、some thinning algorithms use several passes in one thinning iteration. Each pass is an operation to remove boundary pixels from a given direction. Pavlidis 8 and Fiegin and Ben-Yosef 9 have developed effective algorithms using this method. However, both the time complexity and memory requirement wil

19、l increase. In the latter, as the pixels are only flagged, the state of the bitmap at the end of the last iteration will used when deciding which pixel to delete. However, if this flag map is not used to decide whether a current pixel is to be deleted, the information generated from processing the p

20、revious pixels in current iteration will be lost. In certain situations the final skeleton may be badly distorted. For example, a line with two pixels may be completely deleted. Recently, Zhou, Quek and Ng 10 have proposed an algorithm that solves the problem described earlier and is found to perfor

21、m satisfactorily while providing a reasonable computational time. The thinning effect is illustrated in Figure 4Figme 4 Ridgeline Thinningli Before thinningb After thinning3. FEATURES EXTRACTIONThe two basic features extracted from the image are ridge endings and ridge bifurcation. For fingerprint i

22、mages used in automated identification, ridge endings and bifurcation are referred to as minutiae. To determine the location of these features in the fingerprint image, a 3x3 window mask is used (Figure 5). M is the detected point and Xi X8 are its neighboring points in a clockwise direction. If X n

23、is a black pixel, then its response R (n) will be 1 or otherwise it will be 0.If M is an ending, the response of the matrix will beCn Z J?(X +1) 7?(Xr) 1= 2(2)/Iwhere R(9)=R(1). For M to be a bifurcation,8Cn - £ I R«k +1)- R(k) 1= 6(3)/=ifor example, if a bifurcation is encountered during

24、extraction, mask will contain the pixel information such as R(1) = R(3)= R(4) = R(6) = R(7) = 0, R(2) = R(5) = R(8) = R(9) = 1, anda砍4 + 1)A(4)l= 6.A =1For all the minutiae detected in the interpolated thinned image, the coordinates and their minutiae type is save as feature file. At the end of feat

25、ure extraction, a feature record of the fingerprint is formed.4. MATCHINGFingerprint matching is the central part of this paper. The proposed technique is based on structural model of fingerprints 11. One of the major breakthroughs of this method is its ability to mach fingerprints that are shifted,

26、 rotated and stretched. This is achieved by a different matching approach. As it is clear that this algorithm matches the two fingerprint images captured at different time. This matching is based on the minutiae identification and minutiae type matching. Matching procedure is complex due to two main

27、 reasons;1) The minutiae of the fingerprint captured may have different coordinates2) The shape of the fingerprint captured at different time may be different due to stretching.An automated fingerprint identification system that is robust must have following criteria:1) Size of features file must be

28、 small2) Algorithm must be fast and robust3) Algorithm must be rotationally invariant4) Algorithm must be relatively stretch invariantTo achieve these criteria, the structural matching method described by Hrechak and McHugh 11 is adopted as the basis of our recognition algorithm, with changes made t

29、o the algorithm, to provide more reliable and improving overall matching speed. This matching represents the localidentification approach, in which local identified features, their type and orientation is saved in features file, is correlated with the other imageatu ress fextrTbtednlodel is shown in

30、Figure 6.Figure 6 Local Features Structural ModelFor each extracted features on the fingerprint, a neighborhood of some specified radius R about the central feature is defined and then Euclidean distance and relative angles between the central point and the other point is noted with the point' s

31、 type. Since the distance among the pointremains the same throughout the life. So this technique works well for the rotated and shifted images.5. CONCLUSIONA fingerprint recognition algorithm that is fast, accurate and reliable has been successfully implemented. This algorithm can be modified, intro

32、ducing the ridgeline count, and then could be used in online and real time automated identification and recognition system.REFERENCES1 MOAYER, B., and FU, K.S.:'system approach for fingerprint pattern recognitionIEEE Trans.,',1986,PAMT-8 , (3), pp. 376-3872 ISENOR, D.K., and ZAKY, S.G.:'

33、 Fingerprint identification using graph matcPagern Recognit.,',1986,19, (2) pp. 113-1223 TAMURA, H.:' A comparison line thinning algorithms from digital geometry viewpoint' . Proceedingsof fourth international joint conference on Pattern Recognition , Kyoto, Nov. 1978, pp. 715-7194 HILDI

34、TCH, C.J.:' Linear skeleton from square cupboOrdachine Intel., 1969, 4, pp.403-4205 NACCACHE, N.J., and SHINCHAL, R.:' An investigation into the skeletonization approach of HilditchPattern Recognit., 1984, 17, (3), pp. 279-2846 JANG, B.K., and CHIN, P.T.:" Analysis of thinning algorithm

35、s using mathematical morphoEEE'Trans.,Pattern Anal. Machine Intel. , 1990, 12, (6), pp. 541-5517 XU, W., and WANG, C.:' CGT: a fast thinning algorithm implemented on sequential computerEE',8 PAVLIDIS, T.:20,pp133-157Trans.,1987, SMC-17 , (5), pp. 847-851'Algorithm for graphical and i

36、mage processinComput. Graph. Image Process., 1982,9 FEIGIN, G., and BEN YOSEF, N.: pp.108-11210ZHOU, R.W., QUEK, C., and NG, G.S.:1995,16, (12), pp.1267-127511HRECHAK, A.K., and MCHUGH, J.A.:Pattern Recognit., 1990, 23, (8), pp. 893-90412GOUNZALEZ, R.C., and WOOD, R.E.:'Line thinning algoriUhnoc

37、. SPIE ' -,Int. Soc. Opt. Eng., 1984, 397,'Novepsiissjthinning algorithm Pattern Recognit. Lett.,Automated fingerprint recognition using structural matchingDigital image processing ' (Pearson Education, 2002)指纹识别系统摘要指纹识别系统包括图像预处理、特征提取和在个人计算机上进行有效、准确地匹配的功能。图像预处理包括去除噪声、直方图均衡、全局阈值和脊线细化这些必要的

38、特征提取。 提取的 特征，然后存储在指纹匹配的文件里。 这里提出的匹配算法是一种简单、 快速、准确的算法。匹配实验结果准确、 可靠、快速，使用一台个人电脑和指纹识别器实施。提出的指纹算法可 以提供一个有效的自动识别方式，并可以扩展到其他安全或者识别应用。该算法可以进一步植入一个FPGA平台上的个人实时自动识别系统。关键词：生物特征识别；直方图均衡化；山脊细化；脊的结局；脊分岔1、引言指纹识别系统被称为处在生物识别技术的保护伞下。生物识别是指独特的生理（如指纹、人脸、虹膜、视网膜）和行为（如签名、步态）特点，被称为生物识别或简单的生物识别技术，自动识别的个人。在 1893年，它被发现，任何两个人

39、都不会有相同的指纹。在这个发 现之后指纹被应用在刑事鉴定中，直到现在指纹被广泛应用于各种智能识别应用在生活的各 个领域。指纹是人类手指上的图形流的山脊。他们在胎儿发育和指纹配置的大约七个月就完全形成，整个生命中都不会改变，除非由于事故，如瘀伤或指伤。由于不变性和唯一性，指纹识别应用一直是模式识别研究人员和执法机构的极大兴趣。统计方法的指纹特征是提取并存储传统上，指纹识别已经通过任何统计或句法的方法进行。在一个N维的特征向量和决策过程中由一些相似的措施决定。在句法方法，一个模式表示 为一个字符串，树1,或图2的指纹特征或模式基元及其相互关系。然后决策过程就是简 单地句法分析或解析过程。因此适合安

40、全应足够合理的计算本文提出的统计方法。计算机平台上的实验结果证明该方法的有效性， 用于一个相对娇较小的数据库。指纹的预处理进行了修改基本的过滤方法,时间与我们的应用程序的目的。完整的过程方框图在图图1框图2、图像预处理为正确和真实的细节提取，图像质量得到改善，图像预处理，特征提取是必要的， 因为我们不能从原始图像中提取所需的点。首先，任何类型的图像中的噪声目前被删除。顺序统计滤波器用于消除噪声的类型，通常发生在图像采集之后，下面的图像预处理技术应用于提高匹配的指纹图像。2、1直方图均衡化使用这种方法使图像不必要的部分强度轻，以便强调需要所需的部分。图2 (a)显示原始图像和图 2 (b)显示直

41、方图均衡化的图像，其中不连续的小区域被删除。直方图均衡 化，让任意像素的输入和输出电平分别是i和I。然后从0到i(0 < i < 255,0 < k < 255)直方图的累积由C(i)= ±H(k)求得其中H(k)为k层灰色的像素数，即一个区域的直方图，C(i)也被成为累积概率。图2(a)原始图像2、2动态阈值阈值的基本目的是提出所需的对象窗体背景。所有数据点的映射。阈值的结果如图3所示。图2(b)直方图均衡的图像阈值仅仅是灰色的水平，平均灰度水平的图3动态阈值结果2、3山脊细化在特征可以被提取之前，指纹必须变薄或镂空，所以，所有的山脊是一个像素厚。当一个像素

42、作为一个边界像素决定，它直接删除图像3-5或标记，但不会删除，直到整个图像被扫描6-7。在这两种情况下是有缺陷的。在前者，每个边界像素的缺失会改变图像中的对象， 因而影响对称对象。为了克服这个问题，一些细化算法使用几个通行证在一个细化迭代里。每通是一个操作删除从一个给定方向的边界像素。Pavlidis 8 and Fiegin and Ben-Yosef 9已经使用这种方法制定了有效的算法。然而，无论是时间复杂度和内存需求将会增加。在后者，像素只有标记，在最后一次迭代结束时的位图的状态将决定删除哪一个像素时使用。但是，如果不使用这个标志图来决定是否要删除当前图像，从以前的像素在当前迭代处理产生

43、的信息将丢失。在某些情况下，最终的骨架可能会被严重扭曲。例如，有两个像素的线，可 以完全删除。最近，Zhou, Quek and Ng 10提出了一种算法，解决了前面所述的问题，并且 圆满的执行，同时提供了一个合理的计算时间。细化效果如图4所示ab图4山脊细化 a细化前 b细化后3、特征提取从图像中提取的两个基本特征是脊末梢和脊分岔。对于自动识别所使用的指纹图像， 神经末梢和分岔被称为脊的特征点。为了确定这些特征点在指纹图像的位置，使用了一个3x3的矩阵窗口(图5)。M是监测点，X1X 8是在顺时针方向的临近点。如果 Xn是一个黑色 像素，那么与 R (n)对应的就是1,否则就是0。如果M是一

44、个结尾，矩阵响应为8Cn = £ |R(k +1)-R(k) =2k 1其中R(9)=R(1) 。 M是一个分支,8Cn= £ |R(k +1)-R(k) =6(3)k=1例如，如果在提取的过程中遇到了一个分支，掩码将包含的彳t素信息如R(1) = R(3)= R(4)=R(6) = R(7) = 0, R(2) = R(5) = R(8) = R(9) = 1, 结果8Cn = £ |R(k + 1)-R(k) =6k=1插值细化图像中检测到的所有细节，坐标和他们的特征点类型被保存为特征文件。在特征提取的最后，指纹特征记录正在形成。X1X2X3X8MX4X7X6

45、X5图5 3 x 3掩膜特征提取4、匹配指纹匹配时本文的核心内容。提出的技术是基于指纹的结构模型11。这种方法的重大突破之一是它对于移动、 旋转和伸展的指纹相匹配的能力。这是通过不同的匹配方法实现的。因为它是清晰的，该算法匹配两个在不同时间捕获的指纹图像。这种匹配是在特征点识别和特征点类型匹配的基础上的。匹配程序是复杂的，由于两个主要原因：1)捕获指纹的特征点可能有不同坐标2)在不同的时间捕获的指纹形状由于伸展可能会不同强大的指纹自动识别系统必须具备下列条件：1)特征文件的尺寸一定要小2) 算法必须快速和强大3)算法必须是旋转不变的4)算法必须相对伸展不变为了达到这些条件，由Hrechak和M

46、cHugh 11描述的结构匹配方法采用我们书别算法 的基础上，通过算法所做的更改，以提供更可靠，提高整体匹配速度。 此匹配代表局部的识别方法，它们的类型和方位这些局部标识的特征保存在特征文件，与其他图像的特征文件有关。模型如图6所示。图6局部特征的结构模型对于每个提取的指纹特征，临近的一些特定的半径为R的重要特征被定义，然后通过点的类型指出中央点和其他点之间的欧几里得距离和相对角度。由于点之间的距离在整个生命期间保持不变。因此这项技术可以很好的用于转移和旋转的图像。5、结论一个快速、准确和可靠的指纹识别算法已经成功实施。该算法可以修改，引入山脊计数，然后可以应用于网上实时自动识别和识别系统。参考文献1 MOAYER, B., and FU