2009 World Congress on Computer Science and Information Engineering

The Research on Offline Palmprint Identification Jiyi Li, Guangshun Shi, Yan Zheng, Yuanfang Liu Institute of Machine Intelligence, College of Information Technical Science Nankai University, Tianjin, China, 300071 [email protected]; [email protected] in the palmprint for the personal identification and verification[4]. And the characteristics of the minutiae make it the most useful and widely used among the features. In our approach we use two general types of minutia structure. They are ridge terminations (T) and ridge bifurcations (B). A general offline palmprint identification based on minutiae could be composed of several steps or parts. For an input gray-scale palmprint image, the rough feature set could be detected from the original images by the image preprocessing and feature extraction methods. And then the rough feature set is purified by the feature postprocessing methods. The matching methods use the information in the feature sets to evaluate the similarity between two palmprints. We have proposed a series of approaches on palmprint identification in our previous work [6,7,8,9, 10]. Each of them faces to a part or an aspect of the whole offline palmprint identification system, while some of them are more general and could be used not only for offline palmprint. To construct a whole system and provide a complete solution for the offline palmprint identification, this paper chooses the approaches from these works and integrates them into a whole. Section 2 presents the proposed system and the approaches. Section 3 presents an experimental result to show the system performance. And section 4 makes the conclusion.

Abstract This paper summarizes and integrates the approaches in our previous works. It proposes a complete system and provides a complete minutiae-based solution for the offline palmprint identification. The improved image preprocessing approach transforms the grayscale offline palmprint images into skeleton images with the palmprint segmentation, image enhancement and thinning algorithms. The novel matching approach is a multi-phases minutiae matching based on both of the local structure and global feature. To construct a complete identification, this paper also introduces the extraction and postprocessing approaches that extract and purify the feature in this system. The experimental results reveal that the system proposed is effective and efficient for the practical application.

1. Introduction Biometrics has been an important issue in the information society nowadays, and has been widely used in many personal identification and verification applications[1]. Since the good performance on the characteristics of universality, immutability, uniqueness, collectability, acceptability, circumvention and large amount of information[2], palmprint has become a kind of very reliable biometric. While the science of fingerprints recognition has been almost well established[3], the complete theory of palmprint recognition approaches is still being established. Some achievements have been reached in various aspects. Since the differences on the image quality, the image size, the topological property, the available feature types, the sample set scale and so on, the approaches for the online palmprint verification or the fingerprint identification may not be used for offline palmprint identification directly. The system that specifics on the offline palmprint is quite required. Kinds of features, such as the minutiae, singular points, ridges, texture and principal line, could be used 978-0-7695-3507-4/08 $25.00 © 2008 IEEE DOI 10.1109/CSIE.2009.548

2. Offline Palmprint Identification 2.1. System Architecture Fig.1 shows the offline palmprint identification system architecture. First the image preprocessing module transforms the input gray-scale palmprint image into the skeleton image. Then the feature extraction module extracts the raw and rough minutiae set. Because of the image quality and preprocessing approaches used, some true minutiae will be missed while many spurious minutiae will be generated. The feature postprocessing module purifies the rough 587

main orientation percentage, ridge thickness and ridgevalley frequency.

minutiae set, eliminating the spurious minutiae as many as possible while keeping the true minutiae. And finally the output is the corresponding final minutiae set of the palmprint. We add this minutiae set into the palmprint feature database which is constructed by the minutiae sets as well as the palmprint images and personal information. Then we make the palmprint matching with the feature sets in the database according to the transaction requests.

Figure 2. Image Preprocessing Approach The image enhancement approach proposed in [7] enhance the image and transform the gray-scale images into binary images. It is used to strengthen the true feature and remove the noise that may produce spurious feature while retaining the physiological characteristics of palmprint. It consists of four steps: 1. Estimate and modify the orientation field. 2. Remove noises in a gray-scale image. 3. Convert a gray-scale image into a binary image. 4. Remove noises in a binary image. It classifies the noises as interior and exterior noises and designs a group of filters which filter an image according to its orientation field. A binarization method based on the local threshold is used to convert gray-scale images into binary images. The thinning approach transforms the binary images into skeleton images. The traditional thinning approaches could produce many spurious minutiae which are caused by spurs and misconnections between ridges. And the excessive erosion is also a common problem in the thinning approach. Moreover, unit skeleton cannot be produced by some of these approaches. The thinning approach proposed in [8], which improves the Rotation Invariant Thinning Algorithm introduced by Ahmed and Ward (A-W)[5], proposes a series of additional rules to solve the above problems. It adds deleting rules to achieve unit skeleton. It integrates morphological operation into the thinning approach to break misconnection. It combines orientation information with thinning rules to eliminate spurs. It adds amendment step to deal with the termination of the ridge in order to eliminate excessive

Figure 1. The System Architecture for the Offline Palmprint Identification

2.2. Image Preprocessing Fig.2 gives the flow of the image preprocessing approach we use in the system. We obtain the skeleton image with a series of image processing approaches from the original gray-scale image. In this paper, the image processing approach we use are mainly based on the approaches proposed in our previous works [6,7,8]. We divide the image into N*N square blocks. The region of interest (ROI) confirmed by the image segmentation is described by blocks. The approach will also confirm the too blurry blocks (unrecoverable low quality blocks). We just extract the feature in the unblurry blocks in the region of interest. We define them as valid blocks. It could reduce the time cost and spurious feature in the extraction and postprocessing part. The detailed segmentation approach has been proposed in [6]. It uses the white pixel proportion and block variance of the blocks to locate the blank blocks and non-blank blocks. Since the information on the finger is not the object we consider. It detects the blocks that contain the knuckles and fingers. It regards the knuckle-finger blocks as the convex parts of main palm region to detect them. It also detects the too blurry blocks according to the orientation continuity,

588

Bifurcation-Bifurcation, Termination-Termination and Boundary. The Bifurcation-Termination based part could handle the various kinds of spike situations (Fig.3.d,e,f). The Bifurcation-Bifurcation based part could handle various situations of bridge, ladder and island (Fig.3.g,h,i,j,k). The Termination-Termination part could handle various situations of ridge break and wrinkle (Fig.3.l,m,n,o,p,q). And the Boundary based part could handle the boundary effect (Fig.3.r,s,t) around the ROI. For the postprocessing approach, the processing sequence on the processing parts and spurious feature types has an important influence on the overall result. Each processing part or each spurious feature type is not independent with each other absolutely. For example. It is also quite common that several minutiae construct more than one type of spurious feature structure. A more rational sequence is to handle the more reliable processing part in the earlier step. The processing sequence in our system is the above introduced sequence. On the other hand, although there is a processing sequence in the implement, for each part of the postprocessing, the processing object is the whole original rough minutiae set. At the end of each part, the spurious minutiae are not deleted from the memory really, but marked. Finally only the undeleted minutiae are output as the final feature set.

erosion phenomenon. And it introduces the conception of simple rules and complex rules in order to achieve high computational efficiency.

2.3. Feature Extraction and Postprocessing The feature extraction approach extracts both of the palmprint minutiae feature and the local ridge feature, builds the relationship among the feature, and constructs the rough feature set. The feature postprocessing approach proposed is based on the statistical and structural information, combing the information of the minutiae attribute, structural relationship in the minutiae subsets, the local ridge and the local region [9]. To extract both of the minutiae and local ridge feature and create the relationship among the feature, the extraction method has a twice traverse process. The first traverse is based on the whole skeleton image and obtains the basic attributes of the minutiae. The second traverse is based on the rough minutiae set and obtains the local ridge information. The whole feature information is divided by blocks. Each feature is stored in the structure of the block it belongs to. And in each block the feature is stored respectively according to the feature type. In the first traversal, we use the Rutovitz Crossing Number (CN) to detect the coordinate of the minutiae in the skeleton image (Fig.3.a). In the second traversal, for each minutia we have detected in the first traversal, we track its associated ridges along the ridge orientation by pixel. The direction of the minutiae is computed based on associated ridges (Fig.3.b,c). We record lots of information for the minutiae in the feature vector. They are the minutia coordinate and direction, the associated ridges orientation, the sequence of the track list, the track step number, the end point type of track list, and the main branch index of the bifurcation. In the feature postprocessing approach, a series of heuristic rules are proposed to deal with various kinds of spurious feature types. The spurious feature type are the spike, ridge break, wrinkle, island, bridge, ladder, boundary effect and so on. Each spurious type also could be divided into kinds of spurious structures or situations. Actually each spurious feature structure is constructed by several minutiae organized in a particular way, it could also be seen as a minutiae subset generated by a specified relation. For each spurious situation, we propose the corresponding criterion and operating rules. We also divide the postprocessing approach into several parts according to the type and the position of the minutiae which construct the spurious structures. They are the parts based on Bifurcation-Termination,

Figure 3. Feature Extraction and postprocessing Approach

2.4. Feature Matching The matching approach proposed in our previous work [10] could be summarized as the following. For two minutiae set A and B, we construct a local structure for each minutiae Ai and Bj based on current

589

The experimental result showes that the performance is good for the practical application.

minutia and its neighbour minutiae in the polar coordinate. For each Ai we make the local first matching and obtain the Bj candidates sequence based on the similarity of local structure. From all pairs of the Ai and its Bj candidates, we choose the pair which makes the two minutiae sets most similar to be the Benchmarks. Then we make the alignment based on the transform parameters computed from the Benchmarks. At last we make the global second matching and compute the final similarity. The similarity computation formula uses the results in each phase of the matching approach and various kinds of information including the matching number, distance and direction difference. This approach could consider both of the local similarity and global similarity between the two minutiae sets. In this system we also improve the similarity computation formula to normalized the similarity score into [0,1].

4. Conclusion This paper summarizes and integrates the approaches in our previous works, and proposes a complete minutiae-based system and solution for the offline palmprint identification. The experimental result reveals that the system proposed is effective and efficient for the practical application. For the approaches used in the system, the more detailed algorithm descriptoins, discussions and comparisons with others could be referred to our previous works. As the future work, the system considering other kinds of feature such as singular points, ridge feature, and principal line would improve the performance on both of the speed and accuracy. Actually we have reaches several achievements on this aspect. For example, the approach for the location of the special areas has been proposed in [11].

3. Experimental Result We obtain the original inked impression palmprint cards from the Public Security Department, and scan them into the bitmap image format. The specification of the samples is 2400×2400(pixel), 500DPI, 256-level gray-scale offline palmprint image. We have 100 original palmprint image from 98 different palms. From them we obtain 100 original minutiae sets with the proposed system. For each original set, we make 4 additional sets by transforms, including translation, rotation, deletion and insertion. A series of random parameters are used for each set, minutia and transform.

References [1] A.K.Jain, R.Bolle and S.Pankanti, Eds., "Biometrics: Personal Identification in Networked Society", Norwell, MA : Kluwer, 1999. [2] R.Clarke, "Human identification in information systems: Management challenges and public policy issues", Info. Technol. People, 7(4):6-37, 1994. [3] H.C. Lee and R.E. Gaensslen, Eds., "Advance in Fingerprint Technology", New York : Elsevier, 1994. [4] David Zhang, Wai-Kin Kong, Jane You and Michael Wong, "Online palmprint identification", PAMI, 25(9):10411050, 2003. [5] M. Ahmed and R. Ward, “A Rotation Invariant RuleBased Thinning Algorithm for Character Recognition”, PAMI, 24(12):1672-1678, 2002. [6] Yan Zheng, Yuanfang Liu, Guangshun Shi, Jiyi Li and Qingren Wang, "Segmentation of Offline Palmprint", SITIS '07, pp.804-811, 2007. [7] Yan Zheng, GuangShun Shi, Lin Zhang, QingRen Wang and YaJing Zhao, "Research on Offline Palmprint Image Enhancement", ICIP’2007, I541-I544, 2007. [8] Yuanfang Liu, Yan Zheng, Guangshun Shi, Qingren Wang, "Research on Skeletonization of Palmprint Image", SITIS '07, pp.679-686, 2007. [9] Jiyi Li and Guangshun Shi, "A Novel Palmprint Feature Processing Method based on Skeleton Image", SITIS'08, pp.221-228, 2008. [10] Jiyi Li, Guangshun Shi, Qiu Feng and Hongwu Wan, "A New Point Pattern Matching Method for Palmprint", the 2007 International Conference on Artificial Intelligence and Pattern Recognition (AIPR-07), pp.302-306, 2007. [11] Yan Zheng, GuangShun Shi, QingRen Wang, "Location of Special Areas on Palmprint", CISP '08, pp.786-791, 2008.

Figure 4. Receiver Operating Curve At last we obtain 500 minutiae sets and make them matching with others. Then there is 2100 correct pairs (the two feature sets from the same palm) and 247400 incorrect pairs in them. Fig.4 reveals the ROC curve. 590