(a) (b)

F igure 1: (a), Ridge ending; (b), Ridge bifurcation.

Adaptive Approach to F ingerprint Image Enhancement

Li Wang1, Nandita Bhattacharjee2, Gopal Gupta3, Bala Srinivasan4, 1,2,3,4 Clayton School of IT

Monash University

Clayton, Victoria, Australia 3800 +61 3 990 53293

{Li.Wang; Nandita.Bhattacharjee; Gopal.Gupta ; Srini}@monash.edu

A BST R A C T

The purpose of fingerprint image enhancement is to improve the clarity and quality of its local features. The quality of fingerprint image determines the accuracy and reliability of minutia extraction which also determines the accuracy of an automatic fingerprint recognition system. In this paper, we propose an adaptive image pre-processing approach that can significantly improve poor quality images according to their noise level based on contrast stretching, power-law transformation and Gabor filter. The original image smoothing is performed initially by Gaussian filter, and then processed by the proposed adaptive algorithm, and finally the resultant image is filtered by Gabor filter to get the improved binarized image. Experimental results indicate that the proposed approach improves the quality of image and reduces the noise significantly as compared to other fingerprint image pre-processing approaches. Furthermore, the result of Goodness Index (GI) shows that our proposed approach improves the performance by 9% as compared to conventional Gabor filter based approach and is also better than other reported results. Especially, the performance of GI is improved by more than 40% in some poor quality images.

General T erms Algorithms, Design, Experimentation

K eywords Biometrics, Fingerprint recognition, image enhancement

1. IN T R O DU C T I O N

Fingerprint has been used for individual identification for centuries uniqueness and consistency over time [1-2]. Currently, fingerprint recognition is one of the most reliable biometric techniques for personal identification. One major class of fingerprint recognition techniques is based on minutiae detection. Minutiae are major features of fingerprint,

which include: ridge ending, ridge bifurcation, short ridge, island, ridge enclosure, spur, crossover and bridge [1]. Among these minutiae types, ridge bifurcation and ridge ending are considered as the most prominent local characteristics [3]. Figure 1 shows an example of ridge ending and ridge bifurcation. Therefore, extraction of accurate and reliable minutiae is essential in automatic fingerprint recognition system. However, the accuracy of minutiae extraction is heavily depended on the quality of fingerprint images. Detecting genuine minutiae and removing spurious ones in poor quality fingerprint images is a challenge in fingerprint recognition system.

In minutiae based fingerprint recognition techniques, image enhancement plays an important role. Most researchers have focused on enhancing the ridge and valley structure of gray level fingerprint image. The majority of these approaches concentrate on estimating the local ridge orientation and frequency, and then convoluting with orientation and frequency selective filters such as 2-D Gabor filter [3-6] and log-Gabor filter [7]. Some other approaches are based on directional filters such as directional median filter [8-9] and directional Fourier domain filter [10]. Hsieh, Lai and Wang [11] and Hatami et al. [12] developed wavelet based approaches to improve the quality of fingerprint images by decomposing and reconstructing images in wavelet domain. The above approaches improve the clarity of ridge and valley structure of a fingerprint image that is normal and good quality. However, the performance of these approaches reduces on poor quality images with high noise levels. We present an adaptive noise reduction and image enhancement approach for gray level fingerprint images based on contrast stretching and power-law transformation to reduce the noise and improve the image quality. The remaining of this paper is structured as follows: Section 2 presents the proposed adaptive approach for fingerprint image enhancement. Experimental results of proposed approach and conclusions are provided in Section 3 and 4 respectively.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MoMM2010, 8-10 November, 2010, Paris, France. Copyright 2010 ACM 978-1-4503-0440-5/10/11...$10.00.

MoMM2010 Proceedings Mobile Multimedia Security

42

Gaussian filter tp smooth an fingerprint image

Segmentation

Process the image by proposed adaptive

algorithm

Filtered by Gabor filter

The improved binarized image

F igure 2: The flow chart of proposed approach.

.

2. A D APT I V E APPR O A C H T O PR E-PR O C ESS F IN G E RPRIN T I M A G E We propose an adaptive approach to pre-process a fingerprint image in order to improve the quality of the image. The motivation of developing this approach is that qualities of fingerprint images are different, and even within a fingerprint image, the noise level and quality are not the same in its different regions. Therefore, we propose an adaptive approach to enhance a fingerprint image according to its noise and quality level, resulting in improved images as compared to other conventional approaches. Figure 2 shows the steps of our approach:

1. The first step is to use Gaussian filter to smooth the image;

2. This is followed by applying the proposed adaptive approach to enhance the image by removing the noise and increasing the contrast;

3. Next, we use Gabor filter to improve the ridge and valley continuity and clarity.

4. Finally, we binarize the resultant image for future minutiae extraction.

2.1 Gaussian filter to smooth the fingerprint image

(a) (b)

F igure 3: (a): A sample original finger print; (b): the result image after convoluted by Gaussian filter

Gaussian filter is a classical low-pass filter which can smooth an image, so that the edges between ridges and valleys will be smoother especially for dry fingerprints. We use Gaussian filter before the proposed approach in order to smooth the small breaks existing on ridges as our approach may enlarge those small breaks and this problem can be eased by using Gaussian filter. We use 3x3 Gaussian mask in our approach. Figure 3 shows the result of convolution between the Gaussian mask and a sample fingerprint image. Note the fingerprint is from FVC2002

database.We can see that the intensities of ridge lines are more even but ridges and valleys are still as clear as the original image.

2.2 Segmentation

(a) (b)

(c) (d) F igure 4: M ask generation of the image in F igure 3(a). (a): The generated raw mask; (b): The final mask; (c): masked image of F igure 3(a) where the background is set to be white; (d): The boundary of the mask.

Segmentation refers to separate a fingerprint image into foreground and background of which foreground means the recoverable ridge and valley areas of a fingerprint image while background means the unrecoverable, non-ridge and non-valley areas [1]. After segmentation, a mask is generated to shield the background. We calculate gray scale variance and mean intensities of a fingerprint image introduced by [13-14] to generate its mask. A fingerprint image is first divided into 5X5 pixel blocks and for each block we calculate the mean intensity and standard deviation (or variance) (STD). For a mxn block B, the mean intensity M is calculated by equation (1), and then the STD V may be calculated by equation (2).

 

 

Then the mask can be generated as follows:

(3)

T1 and T2 are thresholds selected by experimental observation. The selection of T1 and T2 may be adjusted according to different

Mobile Multimedia Security MoMM2010 Proceedings

43

databases to get the optimized results. In our case, we choose T1 = 0.75*mean STD of all blocks and T2 = mean intensity of the whole image + mean STD of all blocks. Because we choose smaller blocks rather than 16X16 blocks in traditional approaches such as in [3], some foreground areas of an image may be classified as background areas as shown in Figure 4 (a). Therefore, we firstly remove small black blocks inside the raw mask; next we convolute it with a 32X32 size median filter; the final mask is shown in Figure 4(b). Figure 4(c) and (d) show the effect of the mask generated by our approach.

2.3 Proposed algorithm for enhancement of fingerprint images In this part we will firstly introduce two background techniques which are power-law transformation and contrast stretching. Next we describe the proposed algorithm which is designed based on above two techniques. Finally we set rules to estimate parameters in the proposed approach adaptively.

2.3.1 Power-law transformation and contrast stretching Power-law transformation and contrast stretching are common technologies used in digital image processing to improve the contrast and brightness of an image [13]. The equation of power-law transformation is as follows [15]:

(4) Where s is the output gray level, r is the input gray level, c and are positive constants used to control the shape of the transform. When < 1, a narrow range of dark input intensity values map into a wide range of output values. It means that the output image becomes brighter than input image. When >1, the effect of generated curve is opposite to <1. So power-law transformation controls brightness of the output image by selecting different values of . Figure 5 shows the plots of power-law transformation with c = 1 and different values of .In proposed approach, power-law transformation is not applied to the entire input image but parts of the image in a specific range of intensity values, which will be described in next section. Contrast stretching is a linear function which can increase the contrast of an image [13]. Due to its linear property, this transform does not lose data information other than intensity values. Figure 6 shows typical contrast stretching where L is the gray level of an image and is equal to 256, r is the input pixel intensity, s is the output pixel intensity, (r1, s1) and (r2, s2) are two points that control the shape of the transformation [15]. The following equation is used in contrast stretching:

Because the input intensity in the range of [r1, r2] is mapped to a wider range [s1, s2], the contrast of an image is improved. However, two pair of points (r1, s1) and (r2, s2) have to be moderately selected, and their best values depend on different images, which make implementation challenging.

Contrast stretching and power-law transformation have been widely used for adjusting the contrast and brightness in digital image processing [13]. However, we develop a new algorithm based on two approaches above to reduce the noise and enhance fingerprint images. Next section illustrates how this algorithm works.

F igure 5: Plots of power-law transformation with different values of .

F igure 6: Plots of typical contrast stretching.

2.3.2 Proposed algorithm for fingerprint image enhancement Analyzing the histogram of fingerprint images has led to the proposed algorithm. An image histogram is a graphical representation of the tonal distribution in a digital image [13]. Thus, a gray level fingerprint image histogram indicates the distribution of gray levels. Figure 7 shows the histogram of Figure 3(b). We can see that two peaks in this histogram around gray level 40 and 100 along x axis which roughly divide the gray level into 3 parts. The first part can be regarded as gray levels of ridges,

0 L/4 L/2 3L/4 L

L/4

L/2

3L/4

L

=0.1

=0.2 =0.4

=0.67 =1

=1.5 =2.5

=5 =10

Input Gray Levels

Out

put G

ray

Leve

ls

0 L/4 L/2 3L/4 L

L/4

L/2

3L/4

1

Input Gray Levels

Out

put G

ray

Leve

ls

(r1,s

1)

(r2,s

2)

MoMM2010 Proceedings Mobile Multimedia Security

44

F igure 9: The result of our approach of segmented F igure 2(b) with r1 = s1 = 50, r2 =160 and = 1

and the second part as the gray levels of edges of ridges and noise while the last part is the gray levels of valleys and background.

F igure 7: H istogram of F igure 3(b)

Based on the above analysis, we construct a new equation to improve fingerprint images as follows:

Where we apply power-law transformation between (r1, s1) and (r2, s2), and we set s2 = 255 so that the input values bigger than r2 are clipped because we regard this part as valley and background regions. The three parts of equation (6) are applied to three different ranges of input gray levels as discovered above. Figure 8 shows the plots of proposed algorithm with < 1. This algorithm is similar to contrast stretching out we use non-linear transformation in the middle part. That is because contrast stretching can improve the contrast of a fingerprint image, but it does not change the data information except intensity values [13]. Therefore, information of noise is still kept by applying contrast stretching. By applying power-law transformation, the noise level can be reduced better than by linear transformation by selecting according to the noise level. The implementation of our algorithm is shown in the following steps:

1. First, a typical contrast stretching is performed in order to distribute gray levels of an image to [0, 255]. That is because not all fingerprint images occupy the full gray levels of [0,255] as shown in Figure (7). It is performed by applying equation (5) as r1 = min(r), s1 =0, r2 = max(r) and s2 =255, note 5% of the data is firstly saturated at low and high gray levels of the original image in order to remove the interference caused by such small amount of pixels.

2. We then apply the algorithm on the remaining image, block by block and the block size is 16x16. The reason we process the image block by block is that there are different levels of

noise even in one fingerprint image. Therefore, different values of parameters can be selected for different blocks in order to get improved result. The rules of parameter estimation are introduced in the next section.

F igure 8: Plot of equation (6) with < 1.

Figure 9 shows a result of our approach with r1 = s1= 50, r2 =160 and = 1, note this image is segmented to mask the non-ridge and non-valley areas.

2.3.3 Parameters Selection There are a number of undefined parameters in equation (3) which are r1, s1, r2 and in equation (6). For different blocks, the best choice of these parameters may be different. Therefore, a scheme which can automatically estimate the best values for parameters is needed. We use several variables to adjust the values of parameters as shown in Table 1.

Table 1: Variables for parameters estimation

Variables Description Mean intensity(M) Mean of gray levels for each block.

Variance (V) Variance of gray levels for each block.

Reliability of block orientation (R)

The reliability of orientation for each block, high reliability shows the ridge and valley areas are clear, and low reliability shows noise level is high.

0 L/4 L/2 3L/4 L0

L/4

L/2

3L/4

L

Input Gray LevelsO

ut P

ut G

ray

Leve

ls

r2,s2

(r1,s1)

Mobile Multimedia Security MoMM2010 Proceedings

45

F igure 10: The result image of our approach of segmented F igure 3(b) by adaptive parameter selection.

Calculation of mean intensity M and variance V refer to equation (2) and (3).Reliability of orientation indicates the quality of ridge and valley areas [1]. The calculation of reliability R is explained in section 2.3.1. The value R is low for noisy and corrupted regions and high for good quality regions [1]. For low reliability blocks, if the mean intensity M is low then should be set smaller than 1 to increase the mean intensity and contrast; and if the mean intensity M is high then we set bigger than 1 to decrease the mean intensity and increase the contrast. The reliability R generated near singular points are normally low [1], and the mean intensity M also is relatively lower than other regions in the same fingerprint image, so the quality of singular points areas can be improved by setting smaller than 1. Table 2 shows our estimation based on the experiments of different sets of fingerprint database from Fingerprint Verification Competition 2002 (FVC2002) [16]. Figure 10 shows the resultant image of Figure 3 (b) by adaptive parameter selection. We can see that Figure 9 and 10 almost have the same effect.

Table 2: Estimation of parameters in equation (3).

2.4 Gabor F ilter Implementation and Binarization Gabor filter has the features of both orientation and frequency selection [1], so it is especially suitable for fingerprint images. We chose Gabor filter as it can fix small breaks on ridges. Besides, Gabor filter can further remove undesired noises and smooth the ridge and valley structure. The equations for 2-D even symmetric Gabor filter are as follows [1, 17]:

(7)

(9) f and denote the frequency and orientation of local ridges respectively, whereas and represent the standard deviation of Gaussian envelope along x and y axes respectively, and they are both set to be 4.0 based on empirical data [3]. Therefore, before using Gabor filters, two parameters that should be estimated are local ridge orientation and frequency.

2.4.1 Estimation of local ridge orientation and frequency Local ridge frequency and orientation estimation are two important data sets required in subsequent filtering process. Local ridge orientation (i, j) at pixel (i, j) refers to the angle that ridge crossing through a small neighborhood forms with horizontal axis, whereas local ridge frequency at pixel (i, j) refers to the inverse number of ridges per unit length along a hypothetical segment centered at (i, j) and orthogonal to the local ridge orientation (i, j)[1]. In this work, we choose the approach proposed by Kass and Witkin [18] and Bazen and Gerez [19] as they also developed an approach to estimate reliability of orientation at (i, j) from the concordance of different orientation estimated in a neighborhood of (i, j). The steps of estimation are as follows: 1. Divide the input image in size WxW blocks (we use size

16x16); 2. Compute Gx(i,j)and Gy(i,j) which are two gradient

components along x and y direction at pixel (i,j) respectively. They are computed by convoluting with horizontal and vertical Sobel operator and for each block W. We define:

 

3. The orientation of each block centered at (i, j) by:

4. Construct a 1-D wave by projecting the gray levels of all pixels in each block along the direction orthogonal to block orientation.

5. Let T(i, j) be the average number of pixels between consecutive peaks in constructed 1-D wave, then the frequency F(i, j) = 1/T(i, j).

6. The reliability R(i, j) centered at block W is computed as follows:

Parameter Rules of estimation

r2 selection if M>200, r2 = M + V; else if 150< M <=200, r2 = M+ V/2; else if 128< M <=150, r2 = M; else if M<=128, r2 = M- V/2;

r1 and s1 selection

r1 = max(M V, 25); s1 = r1;

selection if M >= 200, = 1.2;

else if R >=0.7, = 1;

else if R <= 0.5, = 0.5;

else if 0.5<R <0.7 and M > 150, = 1.2;

else if 0.5<R <0.7 and 100 <= M <= 150, = 0.8;

else if 0.5<R <0.7 and M < 100, = 0.6;

MoMM2010 Proceedings Mobile Multimedia Security

46

F igure 11: The result image of F igure 10 by Gabor filtering

2.4.2 Result of Gabor filtering and binarization We divide the image into 16x16 blocks and calculate the average orientation and frequency of each block as input to the Gabor filter. Then, we design a Gabor filter bank of 60 filters with angle increment of 3 . For each block, it is convoluted with Gabor filter with closed angle. Figure 11 shows a sample result of image in Figure 10, note the orientation and frequency are calculated based on Figure 3(b).

3. R ESU L TS O F E XPE RI M E N TS We use the database set Db3 of FVC2002 to evaluate the result of our approach. For good quality fingerprint images such as in Figure 3, our approach does not show a large improvement upon the original one. However, when the noise level is high, our proposed approach can significantly reduce the noise level and improve the accuracy of minutiae detection. Figure 12 shows a result comparison between Gabor filtering with and without our adaptive fingerprint image pre-processing approach implemented. The result shows that much of the noise has been removed and ridges and valleys are smoother and clearer, so that the detected minutiae will be more reliable since many spurious minutiae have been removed. More examples are shown in Figure 13.

(a) (b)

(c) (d)

F igure 12: A comparison of fingerprint image pre-processed by our approach and without our approach. (a): The original fingerprint image which is an interclass image of F igure 3 but with higher noise level; (b): The pre-processed image by our

approach; (c): The binary image processed by Gabor filter without pre-processing using our adaptive approach; (d): The binary image processed by our approach.

We also use goodness index (GI) [20] to evaluate the performance of this proposed approach. The GI is defined in [20] as follows:

Where p represents the total number of paired minutiae and t represents the ground truth minutiae in a given fingerprint image, a and b are the number of missed and spurious minutiae respectively, and u is the total number of detected minutiae. We calculate GI by applying our approach to 20 fingerprints in DB3 of FVC2002 containing both good and bad quality images. Table 3 shows the results of GI by comparing our proposed method with the traditional approach. From Table 3, we can see that our proposed approach can detect minutiae more accurate than traditional Gabor filter based approach and improve the performance by 9% on average. Especially, our proposed approach can improve the accuracy of minutiae detection of poor quality fingerprint images such as sample 4 and 12 in which the GI is enhanced by more than 40%. Table 4 shows the comparison between our proposed approach and other approaches. Our average GI is higher than other reported results. Table 3: Comparison between traditional Gabor filter based approach and our proposed approach.

Image Goodness Index

T raditional approach based on Gabor filter

Proposed approach

1 0.60 0.60 2 0.65 0.65 3 0.75 0.67 4 0.29 0.50 5 0.72 0.79 6 0.71 0.71 7 0.79 0.69 8 0.56 0.48 9 0.74 0.72

10 0.60 0.75 11 0.50 0.68 12 0.43 0.83 13 0.42 0.68 14 0.55 0.84 15 0.67 0.72 16 0.49 0.59 17 0.69 0.72 18 0.72 0.78 19 0.40 0.70

Mobile Multimedia Security MoMM2010 Proceedings

47

20 0.74 0.72 STD 0.14 0.09

Average 0.60 0.69

Table 4: Comparison between other approaches and our propsoed appraoch

M ethods M inG I M ax G I Agerage G I

Hong, Wan, Jain [3] 0.29 0.55 0.39 Zhao and Tang [21] 0.18 0.75 0.50

Simon-Zorita et al [22] 0.33 0.76 0.55

Lee and Bhattacharjee[6] 0.31 0.75 0.55

Proposed 0.48   0.84   0.69  

4. C O N C L USI O NS In this paper, we have presented a new segmentation approach and an adaptive approach for reducing the noise and enhance fingerprint images. This adaptive approach is based on improving the contrast and removing unnecessary information such as scars and valueless ridges by adaptive threshold selection. The experimental results show that our proposed approach can effectively improve the quality of fingerprint images especially for poor quality images.

(a) (b) (c)

(d) (e) (f) F igure 13: Two more examples of our approach. (a) and (d) are two bad quality fingerprint images with mask boundaries; (b) and (e) are binary images processed by Gabor filter without pre-processing using our approach; (c) and (f) are binary images processed by our approach.

R E F E R E N C ES

[1] Maltoni, D., Maio, D., Jain, A. K. and Parbhakar, S. Handbook of fingerprint recognition. Springer, New York, 2009.

[2] Lee, H. C. and Gaensslen, R. E. Advances in fingerprint technology. Elsevier, New York, 1991.

[3] Hong, L., Wan, Y. F. and Jain, A. Fingerprint image enhancement: Algorithm and performance evaluation. IEEE Transaction on Pattern Anal, 20, 8 (1998), 777-789.

[4] Greenberg, S., Aladjem, M., Kogan, D. and Dimitrov, I. Fingerprint image enhancement using filtering techniques.

MoMM2010 Proceedings Mobile Multimedia Security

48

15th International Conference on Pattern Recognition, (2000), 322-325.

[5] Jang, W., Park, D., Lee, D. and Kim, S. J. Fingerprint image enhancement based on a half Gabor filter. Advances in Biometrics, Proceedings, (2006), 258-264.

[6] Lee, C. E. and Bhattacharjee, N. Fingerprint Image Processing and Minutiae Extraction for Fuzzy Vault. In Proceedings of the The 7th International Conference on Advances in Mobile Computing & Multimedia (Kuala Lumpur, Malaysia, 2009). Emerald.

[7] Wang, W., Li, J. W., Huang, F. F. and Feng, H. L. Design and implementation of Log-Gabor filter in fingerprint image enhancement. Pattern Recogn Lett, 29, 3 (2008), 301-308.

[8] Wang, S. A. and Guo, Y. C. A Fingerprint Image Enhancement Algorithm Based on Wavelet Transform and Median Filter. Proceedings of the Second International Symposium on Test Automation & Instrumentation, (2008), 719-723.

[9] Wu, C. H., Shi, Z. X. and Govindaraju, V. Fingerprint image enhancement method using directional median filter. Biometric Technology for Human Identification, (2004), 66-75.

[10] Sherlock, B. G., Monro, D. M. and Millard, K. Fingerprint Enhancement by Directional Fourier Filtering. IEEE Image Sign, 141, 2, (1994), 87-94.

[11] Hsieh, C. T., Lai, E. and Wang, Y. C. An effective algorithm for fingerprint image enhancement based on wavelet transform. Pattern Recogn, 36, 2, (2003), 303-312.

[12] Hatami, S., Hosseini, R., Kamarei, M. and Ahmadi, H. Wavelet based fingerprint image enhancement. IEEE International Symosium on Circuit Systems (2005), 4610-4613.

[13] Mehtre, B. M., Murthy, N. N., Kapoor, S. and Chatterjee, B. Segmentation of Fingerprint Images Using the Directional Image. Pattern Recogn, 20, 4, (1987), 429-435.

[14] Mehtre, B. M. and Chatterjee, B. Segmentation of Fingerprint Images - a Composite Method. Pattern Recogn, 22, 4 (1989), 381-385.

[15] Gonzales, R. C. and Woods, R. E. Digital Image Processing. Tom Robbins, 2001.

[16] FVC2002 - F ingerprint Verification Competition, 2002. Available on: http://bias.csr.unibo.it/fvc2002/.

[17] Jain, A. K., Ratha, N. K. and Lakshmanan, S. Object detection using Gabor filters. Pattern Recogn, 30, 2, (1997), 295-309.

[18] Kass, M. and Witkin, A. Analyzing Oriented Patterns. Comput Vision Graph, 37, 3 (1987), 362-385.

[19] Bazen, A. M. and Gerez, S. H. Systematic methods for the computation of the directional fields and singular points of fingerprints. IEEE Transaction on Pattern Analysis and Machine Intelligence, 24, 7 (2002), 905-919.

[20] Chen, Y., Dass, S. C. and Jain, A. K. Fingerprint quality indices for predicting authentication performance. Audio and Video Based Biometric Person Authentication, Proceedings, (2005), 160-170.

[21] Zhao, F. and Tang, X. O. Preprocessing and postprocessing for skeleton-based fingerprint minutiae extraction. Pattern Recogn, 40, 4 (Apr 2007), 1270-1281.

[22] Simon-Zorita, D., Ortega-Garcia, J., Cruz-Llanas, S. and Gonzalez-Rodriguez, J. Minutiae extraction scheme for fingerprint recognition systems. 2001 International Conference on Image Processing, (2001), 254-257.

Mobile Multimedia Security MoMM2010 Proceedings

49