Iris Image Evaluation for Non-cooperative Biometric Iris Recognition System

Iris Image Evaluation for Non-cooperative Biometric Iris Recognition System Juan M. Colores1, Mireya García-Vázquez1, Alejandro Ramírez-Acosta2, and H...
1 downloads 1 Views 844KB Size
Iris Image Evaluation for Non-cooperative Biometric Iris Recognition System Juan M. Colores1, Mireya García-Vázquez1, Alejandro Ramírez-Acosta2, and Héctor Pérez-Meana3 1

Centro de Investigación y Desarrollo de Tecnología Digital (CITEDI-IPN) Avenida del Parque 1310, Tijuana, B.C. México 22510 2 MIRAL R&D, Palm Garden, Imperial Beach, USA 91932 3 Sección de Graduados de Mecánica y Eléctrica (ESIME-IPN), DF., México {colores,mgarciav}@citedi.mx, [email protected], [email protected]

Abstract. During video acquisition of an automatic non-cooperative biometric iris recognition system, not all the iris images obtained from the video sequence are suitable for recognition. Hence, it is important to acquire high quality iris images and quickly identify them in order to eliminate the poor quality ones (mostly defocused images) before the subsequent processing. In this paper, we present the results of a comparative analysis of four methods for iris image quality assessment to select clear images in the video sequence. The goal is to provide a solid analytic ground to underscore the strengths and weaknesses of the most widely implemented methods for iris image quality assessment. The methods are compared based on their robustness to different types of iris images and the computational effort they require. The experiments with the built database (100 videos from MBGC v2) demonstrate that the best performance scores are generated by the kernel proposed by Kang & Park. The FAR and FRR obtained are 1.6% and 2.3% respectively. Keywords: Iris recognition, Kernel, Convolution, Defocus, Quality, Video, MBGC.

1

Introduction

Nowadays, the development of better image quality metrics is an active area of research. The image quality plays a crucial role in the matching system, particularly in automated biometric systems such as iris recognition which performance is based upon matching fine texture information in the annular region between the pupil and the sclera. Some studies report that using a high quality image affects recognition accuracy and can improve system performance [1]. Then, it is necessary to select suitable images with high quality from an input video sequence before the next recognition processing. Otherwise, it can have a negative impact on segmentation algorithms [2,3]. In a recognition system, not all captured iris images are clear and sharp enough for recognition. In fact, in a real capturing iris images system, the person to recognize I. Batyrshin and G. Sidorov (Eds.): MICAI 2011, Part II, LNAI 7095, pp. 499–509, 2011. © Springer-Verlag Berlin Heidelberg 2011

500

J.M. Colores et al.

usually moves his head in diifferent ways gives rise to non-ideal images (with occlusiion, off-angle, motion-blur and defocus) d for recognition. Defocus blur and Motion blur are the major source of iris im mage quality degradation [4]. Defocus blur occurs when the focal point is outside the “d depth of field” of the object to be captured. Depth of fielld is the region of a camera wh hich can capture a well-focused image. It is affected by aperture size, the smaller thee aperture size the greater the depth of field. Motion blur can result either from the relatiive motion of an object or relative motion of the cam mera during exposure time. A sam mple set of all these image problems are shown in figure 1. As it shown, choosing an ap ppropriate image with quality seems a challenge.

Fig. 1. Video sequencees depicting various problems during capturing iris images

To address the problem of o image quality, related work on this subject can be plaaced into two categories [3,5]: lo ocal and global analyses. Local methods try to classify eeach pixel of the iris providing additional a information about each region of the iris textuure. Zhu et al [6], propose a quaantitative quality descriptor by analyzing the coefficientss of particular areas of iris’s textture by employing discrete wavelet decomposition. Chenn et al [2,7], classify iris quality y by measuring the energy of concentric iris bands obtaiined from two dimension waveleets. Ma et al [8], defined a quality descriptor accordingg to characterize out-of-focus an nd motion blur and occlusions. Zhang and Salganicoff [9] examine the sharpness of the t boundary between the pupil and the iris to determ mine defocus in images. Belcherr and Du [10] propose a clarity measure by comparing the sharpness of iris image regions. The major feature of these approaches is that the evaluation of iris image qu uality is reduced to the estimation of a single or a pairr of factors, such as out-of-focuss blur, motion blur, and occlusion. Iris quality should nott be limited to one or two quality factors. Moreover, the majority of previously methhods require involvement of trad ditional segmentation methods that are iterative and tthus computationally expensive. The global methods are quick quality evaluation procedures for selecting the bbest images from a video seq quence and eliminate very poor quality images. Gloobal methods are mainly based d on focus and motion blur estimation. Tenenbaum [[11] proposes a method to determine the focus score using the gradient value of processsed edge image. For checking the motion blur and defocus Jarvis [12], uses the S Sum Modulus Difference. Nayarr [13] adopts the Sum Modified Laplacian to measure the focus score; this method is i based on the absolute values in the second derivattive (Laplacian). However, succh methods target on the entire iris image, and they can generate the wrong focus sccore in case of iris image. Especially, in case of users w with glasses, if the lens is posiitioned for focusing the scratched glasses surface or the

Iris Image Evaluation for Non-cooperative Biometric Iris Recognition System

501

glasses image, such cases may make their focusing scores highest. To overcome this kind of problems, some methods have been proposed to exclusively check the focus score of iris image. In this paper, we present the results of a comparative analysis of four global methods for iris image quality assessment to select clear images in the video sequence. These methods determine defocus and motion blur level in iris images. The goal is to provide a solid analytic ground to underscore the strengths and weaknesses of the most widely implemented methods for iris image quality assessment. The methods are compared based on their robustness to different types of iris images and the computational effort they require. These methods are based on convolution kernels where segmentation is not required because the operator is applied to the entire image, giving the possibility of its implementation at hardware level. This paper is organized as follows. Section 2 explains the principles of kernelbased defocus measurements. Section 3 presents the main representative kernels used for of iris recognition systems. Methodologies for comparing kernels and results are given in Section 4, and Section 5 gives the conclusion.

2

Measurement of Blurring and Defocus in Iris Images

In [14], J. Daugman stated that an effective way to estimate the degree of focus in a broadband image is by measuring its total power in the 2D Fourier domain at higher spatial frequencies, because these are the most attenuated by defocus. Figure 2 shows the total high frequency power in the 2D Fourier spectrum of two iris images assessing the focus of the image. This is a common method applied to image quality assessment [14-17]. In an iris recognition system if an iris image can pass a minimum focus criterion, it will be used for recognition. Thus, it needs a discrete formulation to obtain only the high frequency power.

Fig. 2. a) Clear Image. b) Defocus Image. The total high frequency powers in the 2D Fourier spectrum are 8.772 x 106 and 4.509 x 106 respectively.

502

J.M. Colores et al.

This can be solved filtering the low frequency part of the image, computing the total power in higher frequency bands of the processing image (low frequency filtered image) and setting a predefined threshold [14-17]. 2.1

Filter Low Frequencies in Images

In image processing one of the most common linear operations is filtering. It is known that image filtering can be performed both in the frequency domain or spatial domain. Convolution in the spatial domain is a simple mathematical operation, in which a kernel (matrix) of numbers is multiplied by each pixel and its neighbors in a small region.

Fig. 3. Illustration of applying a convolution kernel to an image in the spatial domain

There are many spatial-domain kernels such as a high-pass filter that can be applied to an image by convolving the kernel with the original image. The figure 3, illustrates graphically the process for a single placement of the kernel. The equation for the two-dimensional discrete convolution is given by: ,

,

where image. 2.2

,

is an image,

,

,

is the convolution kernel, and

(1)

,

is the filtered

Computing the Total Power of an Image

In mathematics, Parseval's theorem shows that the Fourier transform is unitary, i.e., the sum (or integral) of the square of a function equals the sum (or integral) of the square of its transform. The total power of the signal is equal to the total power of its Fourier transform along all its frequency components. The total power, P, in a 2D signal can be measured either in the spatial domain or the frequency domain. For an image , in the spatial domain with the MxN dimension, the total power is

Iris Image Evaluation for Non-cooperative Biometric Iris Recognition System

503

calculated using the Parseval's theorem, so the total power for 2D discrete space signals are calculated by the equation (2). 1

3

,

2

(2)

Convolution Kernels

In order to obtain the total power in higher frequency bands of the image, a proper high-pass convolution kernel is really important. In this section, we will give a brief description of the four convolution kernels most frequently presented in scientific literature to determine the defocus degree in eye-iris images. 3.1

Daugman’s Convolution Kernel

In his pioneering work [14], Daugman proved that the defocus primarily attenuates high spatial frequencies. In this band-pass filter, the central frequency is around 0.28125 with a bandwidth of 0.1875 in which the attenuation is less than 3db with respect to the central frequency. The u , v , 2-D Fourier transform is represented as equation (3) and its spectrum response are shown in figure 4b. ,

sin

sin

sin 2 sin 2 4

(3)

Due to this relationship, he improved the convolution operation in real-time, id est., to reduce the computational complexity of the Fourier transform he proposed a high pass 8×8 convolution kernel to extract the high frequency of an image. The weights consist of two square box functions, one of size 8x8 with amplitude -1, and the other one of size 4x4 and amplitude of +4.

Fig. 4. a) The 8x8 Convolution kernel proposed by Daugman b) The frequency response (Fourier Spectrum)

504

3.2

J.M. Colores et al.

The Convolution Kernel of Wei et al.

Wei et al. [15] also suggest a convolution kernel, with a similar shape as Daugman’s for detecting defocused on still and segmented images. Additionally, they detect other problems presented in iris images; motion blur and occlusion. Each problem has its own peculiarity, so, the three features are used to classify them using Support Vector Machine (SVM) that is a machine learning algorithm method used for classification [18]. The total quality of an image according with their method is a vector Q (q1, q2, q3), where the values represent the levels from defocus, motion blur and occlusion respectively. To determine the defocus degree they proposed a 5×5 convolution kernel as shown in figure 5. Compared with Daugman’s 8×8 convolution kernel is also a lower frequencies filter but computationally less demanding. The operator is formed by three box functions, one of size 5x5 with amplitude -1, one of size 3x3 with amplitude +3, and the last one of size 1x1 with amplitude -2.

Fig. 5. a) The 5x5 Convolution kernel proposed by Wei et al. b) The frequency response (Fourier Spectrum)

The kernel Fourier spectrum is shown in figure 5b. It is a band-pass filter with central frequency around 0.4375 and bandwidth of 0.3125 in which the attenuation is less than 3db with respect to the central frequency. 3.3

Laplacian of Gaussian Convolution Kernel

J. Wang et al. [16] propose a convolution kernel operator based on a Laplacian of Gaussian function (LoG). The Laplacian gives a 2-D isotropic measure of the second spatial derivative of an image. In the image the high intensities regions represent the rapid intensity change, which are used for edge detection. The Gaussian smooth filter is first applied to the image to reduce the noise sensibility of the second derivative. Laplacian x , y of an image with pixel intensity values x , y is given as follows. ,

(4)

Iris Image Evaluation for Non-cooperative Biometric Iris Recognition System

505

The combination of both filters gives a function centered on zero with standard deviation σ as is given by the equation 5: 1

,

1

2

(5)

Since the image is represented as a set of pixels, the authors sought a discrete convolution kernel that can approximate the Laplacian operator. They set different values of the Gaussian. Finally they used σ 1.4 that produce a LoG operator as shown in figure 6.

Fig. 6. a) The 9x9 Convolution kernel based in Laplacian and Gaussian filter b) The frequency response (Fourier Spectrum)

3.4

The Convolution Kernel of Kang and Park

Kang & Park [17] propose 5x5 pixels sized convolution kernel as shown in figure 7. It is a band-pass filter and its central frequency is around 0.2144, with a bandwidth of 0.6076 in which the attenuation is less than 3 db with respect to the central frequency. The u, v in the 2-D Fourier domain is plotted en figure 7b and represented as equation (6). ,

sin

3 2 9 4

sin

3 2

sin

5 5 sin 2 2 25 4



sin

1 2 1 4

sin

1 2

(6)

The kernel consists of three square box functions, one of size 5x5 with amplitude -1, one of size 3x3 and amplitude +5, and other of size 1x1 and amplitude -5 (see figure 7). They argue that their 5x5 pixels convolution kernel contains more high frequency bands than the 8x8 pixels convolution kernel proposed by Daugman [14]. According to the authors the operator can detect much better the high frequency of iris texture, using less processing time due to the short sized kernel.

506

J.M. Colores et al.

Fig. 7. a) The 5x5 Convolution kernel proposed by Kang & Park b) The frequency response (Fourier Spectrum)

4

Experimental Results

4.1

The Iris Video Dataset

In order to evaluate the performance of the four methods for iris image quality assessment, we selected the Multiple Biometrics Grand Challenge “MBGC.v2” database [19]. It was collected during the spring of 2008 by The Computer Vision Research Lab at the University of Notre Dame and provided 986 near infrared eye videos. All videos were acquired using an LG2200 EOU iris capture system [20] (see figure 8a). The camera uses near-infrared illumination of the eye. The iris video sequences were digitized by a DayStar XLR8 USB video digitizer attached to a Macintosh host system and stored in MPEG-4 format. The size for each frame in the video has 480 rows and 640 columns in 8 bits-grayscale space (intensity values between 0 to 255). The MBGC database presents noise factors, especially those relative to reflections, contrast, luminosity, eyelid and eyelash iris obstruction and focus characteristics. These facts make it the most appropriate for the objectives of our work. We produced our own database of 4432 iris images from the MBGC iris video database v2. 100 videos were randomly selected from this database. Our database contains 2077 clear iris images (positive samples) and 2355 defocused iris images (negative samples). The all 4432 iris images were manually checked and selected in positive and negative samples by a subjective process (based on human perception of defocus). In iris images where it had uncertainty about whether it was a clear iris image or a defocused one, verification tests were implemented. Thus, the negative samples come from those iris images that cannot be segmented by Libor Masek recognition algorithm [21]. The iris images that did not have good iris segmentation were replaced. For instead, iris images that show blinks and offangle, see figure 8.

Iris Image Evaluation for Non-cooperative Biometric Iris Recognition System

507

Fig. 8. a) LG EOU2200 system b) The iris images that do not include enough iris information

4.2

Best Iris Image Selection

The table 1, shows the obtained results, the first column identifies the evaluated kernel, the second and third column show the mean and the standard deviation of the total power at high frequency bands processing each kernel with clear iris images database. The next two columns give the mean and the standard deviation processing each kernel with the defocused iris images database. The first last column indicates the optima’s threshold for discrimination of defocus iris images. We use the receiver operation characteristic (ROC) curves to obtain the optimal decision threshold. Table 1. Results of the evaluated kernel methods

Kernel Daugman Wei et al. LoG Kang & Park

Defocus iris images

Mean

Standard deviation

Mean

Standard deviation

Optimal Decision threshold

55.063 19.928 108.49 25.792

8.9134 4.5584 10.651 6.0658

26.838 12.291 72.88 13.25

5.7642 5.3759 14.658 3.9636

39.9490 14.1631 92.0776 15.8247

Clear iris images

The error percentages for every kernel are shown in the table 2. If an accepted/positive iris image (focus image) is a defocus image, it is called a false accept. The percentage of false accepts is called false accept rate (FAR). If a rejected iris image (defocus image) is a focus image, it is called a false reject. The percentage of false reject is called false reject rate (FRR). The best performance scores (minimum FAR and FRR) was generated by the kernel proposed by Kang & Park followed by the Daugman kernel who presented also low error rates. The worse results performance kernels were presented by kernels proposed in [15,16], these kernels had the highest error rates.

508

J.M. Colores et al. Table 2. Error percentages of evaluated kernel methods

Kernel

FAR (%)

FRR (%)

Daugman Wei et al.

2.8 8.3

3.6 2.6

LoG Kang & Park

3.7

5.2

1.6

2.3

To compare the convolution kernels in terms of speed, we compute the total multiplication count (TMC) [14]. For the Daugman’s convolution kernel: TMC = (8x8x640/4 x 480/4) = 1228800. For the Laplacian of Gaussian Convolution kernel: TMC = (9x9x640/4 x 480/4) = 1555200. For the convolution kernel of Wei et al and convolution kernel of Kang & Park: TMC = (5x5x640/3 x 480/3) = 853333. With this, it is shown that the last two convolution kernels are 30.56% faster than the Daugman’s convolution kernel.

5

Conclusions

In this paper, we present the results of a comparative analysis of four representative convolution kernels for iris image quality assessment to select clear images in the video sequence. The defocus and motion blur assessment allow distinguishing between the images with high quality (clear images) and those with low quality (blurred images). To distinguish between both images classes, an optimal threshold was established. It was experimentally obtained analyzing the database presented in section 4.1. We used the ROC curves to obtain the optimal decision threshold. From the experimental results, it was concluded that the Kang & Park convolution kernel was superior to the other three kernels in terms of speed and accuracy. The results obtained by this kernel showed the lowest error rates (FAR=1.6% and FRR=2.3%). The results suggest that, to add a quality assessment stage to the video iris recognition system could increase the system performance. Acknowledgment. This work was supported by SIP2011 grants from IPN.

References 1. Gamassi, M., Lazzaroni, M., Misino, M., Piuri, V.: Quality assessment of biometric systems: a comprehensive perspective based on accuracy and performance measurement. IEEE Transactions on Instrumentation and Measurement 54, 1489–1496 (2005) 2. Chen, Y., Dass, S.C., Jain, A.K.: Localized Iris Image Quality Using 2-D Wavelets. In: Zhang, D., Jain, A.K. (eds.) ICB 2005. LNCS, vol. 3832, pp. 373–381. Springer, Heidelberg (2005) 3. Kalka, N.D., Zuo, J., Schmid, N.A., Cukic, B.: Image quality assessment for iris biometric. In: SPIE: Biometric Technology for Human Identification III, vol. 6202, pp. D1–D11 (2006)

Iris Image Evaluation for Non-cooperative Biometric Iris Recognition System

509

4. Cao, Y., Wang, Z., Lv, Y.: Genetic Algorithm Based Parameter Identification of Defocused Image. In: ICCCSIT 2008: International Conference on Computer Science and Information Technology, pp. 439–442 (September 2008) 5. Zuo, J., Schmid, N.A.: Global and local quality measures for NIR iris video. cvprw. In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, cvprw, pp. 120–125 (2009) 6. Zhu, X.-D., Liu, Y.-N., Ming, X.: A quality evaluation method of iris images sequence based on wavelet coefficients in region of interest. In: CIT 2004: Proceedings of the The Fourth International Conference on Computer and Information Technology, pp. 24–27. IEEE Computer Society, Washington, DC, USA (2004) 7. Chen, J., Hu, G.S., Xu, J.: Iris: Image Quality Evaluation Method Based on Wavelet Packet Decomposition. Journal oj Tsinghua University (Sci&Tech), 43(3), 377–380 (2003) 8. Ma, L., Tan, T., Wang, Y.: Personal identification based on iris texture analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence 25(12), 1519–1533 (2003) 9. Zhang, G., Salganicoff, M.: Method of measuring the focus of close-up image of eyes. Tech. Rep. 5953440, United States Patent (1999) 10. Belcher, C., Du, Y.: Information distance based selective feature clarity measure for iris recognition. In: Proc. SPIE, vol. 8, pp. 64940E1–64940E12 (2007) 11. Tenenbaum, J.M.: Accommodation in computer vision. Ph.D. thesis, Stanford University (1970) 12. Jarvis, R.A.: Focus optimization criteria for computer image processing. Microscope 24(2), 163–180 (1976) 13. Nayar, S.K., Nakagawa, Y.: Shape from focus. IEEE Trans. Pattern Analysis and Machine Intelligence 16, 824–831 (1994) 14. Daugman, J.G.: How Iris Recognition Works. IEEE Trans. on Circuits and Systems for Video Technology 14(1), 21–30 (2004) 15. Wei, Z., Tan, T., Sun, Z., Cui, J.: Robust and Fast Assessment of Iris Image Quality. In: Zhang, D., Jain, A.K. (eds.) ICB 2005. LNCS, vol. 3832, pp. 464–471. Springer, Heidelberg (2005) 16. Jing, W., He, X.F., Pengfei, S.: An Iris Image Quality Assessment Method Based on Laplacian of Gaussian Operation. In: MVA 2007 IAPR Conference on Machine Vision Applications, Tokyo, Japan, May 16-18, pp. 248–251 (2007) 17. Kang, B.J., Park, K.R.: A Study on Iris Image Restoration. In: Kanade, T., Jain, A., Ratha, N.K. (eds.) AVBPA 2005. LNCS, vol. 3546, pp. 31–40. Springer, Heidelberg (2005) 18. Burges, J.C.: A Tutorial on Support Vector Machines for Pattern Recognition. Data Mining and Knowledge Discovery 2, 121–167 (1998) 19. Multiple Biometric Grand Challenge, http://face.nist.gov/mbgc/ 20. Bowyer, K.W., Hollingsworth, K., Flynn, P.J.: Image Understanding for Iris Biometrics: a Survey. Computer Vision and Image Understanding 110(2), 281–307 (2008) 21. Masek, L.: Recognition of human iris patterns for biometric identification. Master’s thesis, University of Western Australia (2003)

Suggest Documents