A Robust and Real-time Multi-feature Amalgamation Algorithm for Fingerprint Segmentation Sen Wang Institute of Automation Chinese Academ of Sciences P.O.Bo 78 Beiing P.R.China100080 Yang Sheng Wang Institute of Automation Chinese Academof Sciences P.O.Bo 78 Beiing P.R.China100080 Abstract: A critical step in automatic fingerprint identification is to automaticall and reliabl decide hich part of the fingerprint image belongs to the foreground the nois area and the background. In this paper e define the background as an area ecept foreground including nois area blurred area and blank area. In order to ensure that the performance of an automatic fingerprint identification sstem ill be robust ith respect to the qualit of input fingerprint images it is essential to have a good algorithm. In this paper a robust and real-time algorithm for of fingerprints is presented. With almost no cost of time ne features are etracted. The method uses to ne features hich are contrast and main energ ratio and other to features of blocks of piels: coherence and variance. It also uses k-means classifier and 3-nearest neighbor to classif patterns and distinguish hich part of fingerprints is foreground or background. The post-processing reduces the error classification. perimental results sho a significant improvement in fingerprint performance. Also the time required for the identification sstem is reduced. Ke-ords: Biometrics Image processing Fingerprint Segmentation K-nearest neighbor K-means classifier. 1. Introduction Fingerprint identification is one of the most important biometric technologies since the unchangeable fingerprints during human life span and the uniqueness of each individual s fingerprints. A fingerprint is the pattern of ridges and valles on the surface of a fingertip. The uniqueness of the fingerprint can be identified b the characteristics and relationships of bifurcations and endings in ridge or valle (Figure 1. In order to compare to fingerprints a set of invariant and discriminating features are etracted from fingerprint image. Most verification sstems providing a high securit are base on minutiae matching.
used to decide hich part belongs to foreground and hich one to background. Then e present k-means classifier and 3-nearest neighbor to classif patterns (Section 3. Finall e present post-processing and some eperimental results (Section 4. Figure 1: amples of minutiae. Because more poerful and intelligent image processing techniques are possible b development in computer technolog the automatic fingerprint identification sstem appears [1]. Unfortunatel noise image deficienc and deformation ma make reliable minutiae detection ver difficult. Therefore one important step in automatic fingerprint identification is the of fingerprint images. The main purpose of is to determine hich part of fingerprint is foreground and hich one is background. Foreground is the good pattern of ridges and valles. The nois area the blurred area and the blank area in fingerprint are background. Because e ma etract false minutiae in background this can heavil influences the performances of the minutiae etraction process and the performance of the overall fingerprint identification sstem. So the can reduce the number of false minutiae. In man algorithms features the etract can not completel present the feature of piels so their algorithm can onl identif the blank area but can not classif beteen foreground and nois or blurred area. [3] In this paper e present to ne features hich are contrast and main energ ratio (Section. The ne features and the other to features are. The Presentation of To Ne Features.1 Features traction The first step of is to select features. Three features are used in [3] hich are coherence the local mean and the variance. Hoever those three features can not completel present the feature of piel so e select other to ne features. Since fingerprint mainl consists of parallel structures the coherence in foreground is higher than that in background. The block is and the definition of coherence follos: i+ + = G ( u v G ( u v O u= i v= i+ + = G ( u v G ( u v ( O u= i v= i+ + ( = G ( u v + G ( u v O Coherence = u= i v= O O + O ( i G ( i G ( i gradients at each piel.. are the.
. Contrast one of the ne to features One ne feature is the contrast hich is the normalization of variance. The contrast defined as: I( i i m = v = ( I( i m Contrast = m v. The v is variance in one block; the m is mean of one block. The block is and is 16. the image. I is the intensit of.3 Main nerg Ratio the other ne feature The other ne feature is the main energ ratio. A good fingerprint patterns contain narro ridges separated b narro valles and this ridges follo almost parallel to each other. The Fourier spectrum of this small area reveals to high peaks ecept DC components. Hoever hile the fingerprint pattern is not good enough the peaks are not distinct. [4] (Figure. So this method can distinguish fingerprint signal area and noise area. This feature etraction algorithm follos: 1. Divide image to blocks of sizes (16 16;. ach block applies FFT to transform image from spatial domains into frequenc domains. 3. Search the to peaks ecept DC components. The to peaks are smmetrical to the DC components. 4. Because of noise e define main energ ratio as: p1 + p Main nerg Ratio= ( + p1 = F if F( i > p1 30%. s is the value of image in frequenc domains after FFT. The s is circle hich radius is r. The r is the distance from the peak to DC component and is round to the nearest integer. p p1 F is the value of one peak. is the value of the other peak. p is approimatel equal to. p1 p (1 ( (3 (4 Figure : (1 (3 are original images. ( (4 are FFT of (1 (3 respectivel. In the singular point area because the orientations of ridges and valles change so rapidl and the main energ ratio and coherence can not present the feature of piels in this area ver ell the to features can not be used in this area. Hoever e can identif this area [5] and use the other to features to process this area.
3. Design the Classifier In order to be real-time e must minimize the computational compleit. To find the optimal decision boundaries e use k-means classifier and 3-nearest neighbor to classif foreground and background. We select man tpical blocks in both foreground and background. We manuall point out and compute the four features for these blocks. We get about 3000 four-dimensional patterns and put those 3000 patterns into k-means classifier. Before put these patterns into classifier e normalized each pattern first. perimentation shos that identi those patterns into 1 clusters is the best. 8 of these clusters are in foreground and 4 clusters are in background. Outline of the K-means clustering algorithm e used is as follos: 1. Initialize each cluster centers ( l = 0 and normalize each pattern. ( ( (} { z l z l... z. 1 1 l z i ( l is the cluster center in the l times.. Classif 3000 patterns. Through this step each pattern ( p corresponding to each cluster as folloing: if ( p S ( l ( p ( p z ( l < z ( l ( i = 13...1. i S ( l J = ( p S ( l ( = 13...1 ( p z ( l + 1 So the center of each cluster is computed as follos: ( p fing z ( l + 1 = i is is all members in. cluster. 3. Compute the ne cluster center. In step e get all members in the ne cluster and e ill compute the ne center in each cluster to make smallest the sum of the distance from each member to the ne center. We epect minimize J : patterns in. N S 1 N ( p S ( l is the number of 4. Check convergence. In step 3 the ne centers do not change an more. z ( l +1 z ( l ( = 13...1 = If the ne centers do not change an more go to end else go to step. When e finish classifing those patterns into 1 clusters e use a 3-nearest neighbor to classif each block in an input fingerprint image into foreground and background. If a block is foreground then B(i =1 else B(i = 0. After image B is obtained the classification is completed. We can also compute the percentage of foreground. If the percentage of foreground regions is smaller than a threshold then the input fingerprint image is reected.
4. Post-Processing and perimental Results When e finish the classification there ma be some noise. For instance spurious blocks hich belong to one class ma appear in other class. We use Gauss-smooth to reduce noise. It reduces the number of false classification. After the Gauss-smooth operates e get a good image of fingerprints. We present some eperimental result of the algorithm. In our eperiments e avoided using small samples so this algorithm as tested on man different fingerprint images from standard database. We compared this algorithm ith global and local threshold algorithms hich onl used three features: mean variance and coherence. We used k-means cluster hich as trained b 3000 patterns e selected in specific database. We also selected 600 fingerprint images at random in man different standard databases. We divided those images into 194400 blocks and got 194400 patterns from those images. We used our algorithm and other algorithms to distinguish hich block is foreground or background respectivel. Our algorithm classified 7 false blocks in foreground and 15 false blocks in background. When e used global or local threshold the error rate as higher. From Table 1 the result shos that our algorithm is better than the others. We also applied these three algorithms to 00 fingerprint images of size 300 300 captured b a COMS sensor. We divided TM Veridicom each image into 18 18 blocks and used these three algorithms to classif these blocks. We counted the number of the blocks hich are in foreground nois area and blurred area and blank area b human reorganization our algorithm and the other to algorithms respectivel. From result Table our algorithm is better especiall in nois area and blurred area. In nois area and blurred area main energ ratio is smaller than that in foreground hoever mean and variance almost is equal to those in foreground. Coherence in nois area and blurred area is higher than that in foreground but it quite depends on gradient and orientation hich are difficult to accuratel count in each piel. So our algorithm is better than the other to algorithm since e present the to ne features. Figure 3 shos one results of the fingerprint images in 00 images hich are used b different algorithms.
(1 ( (3 (4 Figure 3: (1 is the orient image ( is the result of the global threshold (3 is the result of the local threshold (4 is the result of our algorithm. Black area is background hich is signed b different algorithms. Table 1: Results of different algorithm Our algorithm Local threshold Global threshold Human reorganization Our algorithm Local threshold Global threshold number of false blocks in foreground number of false blocks in background 7 15 1306 6884 1564 1034 error rate.04% 9.87 % 13.17 % Table : Results of different algorithm number of blocks in number of background blocks in error rate foreground In nois area and blurred area In blank area 45309 7386 1105 45914 763 1163 1.86% 48385 4693 117 9.49% 50873 311 10806 17.17%
5. Conclusion and Discussion In this paper e discuss a robust and real-time multi-feature amalgamation algorithm for fingerprint hich should be incorporated in the preprocessing of an automatic fingerprint identification sstem. Since e present to ne features our algorithm largel reduces the error of classification especiall in nois area and blurred area. The to ne features can eactl reflect the difference beteen foreground and background so the make the classification more accuratel. We use K-means classifier and 3-nearest neighbor to classif each block of input fingerprint image into foreground and background. Through eperimental result and human inspection the proposed method provides a real-time and accurate result. We ill etract more efficient features in the future hich can make classification more precise. We can also choice ne classification algorithm for eample neural netork. We think these ma improve the result. Fingerprint Segmentation I Trans. 1996 [3] Asker M. Bazen Sabih H. Gerez Segmentation of Fingerprint Images ProRISC 001 Workshop on Circuits Sstems and Signal Processing 001 [4] Toshio Kamei Masanori Mizoguchi Image Filter Design For Fingerprint nhancement I Trans. 1995 [5] Weiei zhang Yangsheng Wang Singular Point Detection in Fingerprint Image Accept b ACCV 00. [6] L. Hong Y. Wan and A. Jain Fingerprint Image nhancement: Algorithm and Performance valuation I Trans. Pattern Analsis and Machine Intelligence Vol. 0 no. 8 pp.777-789 1998 [7] Zs. M. Kovacs-Vana R. Rovatti M. Frazzoni Fingerprint Ridge Distance Computation Methodologies Pattern Recognition Vol33 pp. 69-80 000 [8] Marius Tico Pauli Kuosmanen A Topographic Method For Fingerprint Segmentation I Trans. 1999 [9] B. G. Sherlock D. M. Monro K. Millard Fingerprint nhancement B Directional Fourier Filtering I Proc. Vis Image Signal Process Vol 141 No 1994 Reference [1] A. Jain L. Hong and R. Bolle On-Line Fingerprint Verification I Trans. Pattern Analsis and Machine Intelligence Vol. 19 no. 4 pp.30-314 1997 [] M. Hassan Ghassemian A Robust On-Line Restoration Algorithm For