nonintrusive iris image extraction for iris recognition-based biometric

0 downloads 0 Views 257KB Size Report
Key words: Iris recognition, biometric identification, morphology, thresholding, hue- ..... With nonoptimized code implemented in MATLAB, it takes 1 second to ...
C IRCUITS S YSTEMS S IGNAL P ROCESSING VOL . 25, N O . 3, 2006, P P. 405–419

c Birkh¨auser Boston (2006)  DOI: 10.1007/s00034-005-0305-6

N ONINTRUSIVE I RIS I MAGE E XTRACTION FOR I RIS R ECOGNITION -BASED B IOMETRIC I DENTIFICATION * Sarp Ert¨urk1

Abstract. This paper proposes a new approach to obtain iris images without requiring the person to look directly into the camera. Most biometric identification methods using iris recognition assume that the eye image of the person is available, and are only concerned with the extraction of the iris from the eye image. These methods require the person to look directly into the camera, which is a rather uncomfortable process. In this paper, a robust approach is proposed to initially locate eye regions within facial images taken from a distance. Then the iris image is extracted from the detected regions. Hence, a nonintrusive and more comfortable iris imaging approach has been implemented. Key words: Iris recognition, biometric identification, morphology, thresholding, huesaturation-intensity space, circle fitting.

1. Introduction Reliable identification systems are required for many applications such as ATM banking, restricted area access control, database access, computer login, building entry, and airport security. Biometrics has been proposed as a reliable source for identification systems. Various biometric features such as facial shape, hand shape, fingerprint, sound characteristics, and iris recognition have been proposed for human identification. Iris recognition stands out as a promising method for obtaining fully automated, secure, reliable, fast, and uncomplicated identification systems. Furthermore, small oscillations occurring in the human pupil cause the iris to constantly display small size changes, which can be used to detect erroneous iris patterns (for example, an iris picture), so as to avoid system penetration. Automated iris recognition was first proposed by Flom and Safir in 1987 [5], but there is no evidence that a working system was implemented at that time. ∗ Received March 5, 2005; revised July 3, 2005; This work was supported by the Turkish Technical

and Scientific Research Council, TUBITAK, under grant EEEAG/103E018. 1 Electronics and Telecommunications Engineering Department, University of Kocaeli, 41040,

Turkey. E-mail: [email protected]

406

Author: In ref 16 date is 1994 not 1996

¨ E RT URK

Daugman implemented the first prototype of an operating iris recognition system at Cambridge University in 1993 [17]. Daugman proposed to utilize circular zones for iris alignment, and two-dimensional Gabor filters for multiscale representation of the iris pattern, using the filter parameters to construct a codeword which is then compared using a Hamming distance criterion for a similarity measure. Later in 1994, an alternative prototype that in many points resembles the Daugman system has been reported by Wildes et al. in [16]. Wildes et al. used a Laplacian of Gaussian filters that split the iris pattern into frequency bands in the form of a Laplacian pyramid, and used a correlation measure computed for all pyramid levels to determine the similarity between acquired and stored patterns. In [1] and [2] it has been proposed to start with the eye image and locate the iris using edge detectors sensitive to circular features, and then the zero crossings of the wavelet transform are used to extract iris information. In [14] the goal is to determine the minimum template size required for iris recognition through the investigation of the biological features of the iris. The approach again starts from the eye image and uses circular edge detection to locate the iris and Gabor filtering-based analysis. In [13] a system is proposed that automatically locates the eye region of a person staying at a distance and then captures the eye image, to enable a broader utilization of iris recognition. However, the system consists of three cameras. While the eye region is identified using two wide angle cameras operated in stereo mode, a third camera is used to capture the eye image. Therefore, the utilized system and approach are rather complex. In [19] it has been proposed to start from the eye image and locate the center of the iris using thresholding, after which the outline of the iris is identified using the maxima in circular grey level differences. After the iris is extracted, the iris image is converted into a rectangular shape, and a local histogram equalization is utilized to compensate for contrast and illumination differences. It has been proposed to utilize neural networks to locate the iris region in [6]. For this purpose, the image is divided into blocks of size 20 × 20 pixels, and the neural network decides whether each block is part of the iris or not. In order to gain robustness against rotation, rotated versions of the iris images are also used in the training of the neural network. It has been proposed to detect reflections and eyelashes in [9] making use of Gabor filters, variance computation of small windows, and a statistical test for uniformity. It has been proposed in [10] to start with the eye image and extract the iris using Canny edge detection together with a homocentric circle approach. In [4] it has been proposed to locate the nose region within facial images and conduct a search from the nose region towards possible eye locations to detect the iris within facial images. Circular edges are used to detect the iris in eye images in the work proposed in [12] and symmetric circular filters are used to obtain iris feature vectors, while the nearest feature line is used to match feature vectors for identification.

N ONINTRUSIVE I RIS I MAGE E XTRACTION

407

Independent component analysis is used in conjunction with competitive learning in [7] for iris recognition. The approach starts from an image of the eye region, edge detection is applied, and the iris is detected based on roundness. Separability filters are used in [8] to detect valleys in facial images and extract the iris locations using the Hough transform. Iris recognition based on comparing the fractional Fourier transform of iris images using characteristic matrices is proposed in [18]. In [15] the iris is located in a facial image using the Hough transform, and each pixel is assigned a vote that is proportional to the gradient at that point. A multiresolution approach is proposed in [3] to locate the iris in eye images using gradient force and uniformity measures. In [11], a frequency distribution-based approach is proposed to identify the quality of an iris image. Most iris recognition techniques assume that the eye image is available and it is only required to locate the iris within the eye image. This is typically possible if the user looks directly into the camera; however, this is quite uncomfortable for the user and rather intrusive. Methods that have been proposed to identify the iris location from facial images captured from a distance either require a complex system [13] or include a complex algorithm that is usually time-consuming and therefore not appropriate for near real-time iris recognition systems [6], [4], [8], [15]. A simple approach to initially locate the eye region in facial images taken from a distance is proposed in this paper to provide a fast and nonintrusive method for obtaining the eye images. Furthermore, a relatively fast technique to extract the iris from eye images is presented so that the iris pattern can be extracted reasonably fast without causing discomfort to the user. The proposed approach is novel in that it makes use of the hue-saturation-intensity (HSI) color space to identify the eye region by locating the white parts of the eye and utilizes shape knowledge to correctly identify the eye regions. Furthermore, the iris is extracted using a simple approach consisting of thresholding, edge detection, and circle fitting. Hence, a reasonably fast approach to extract the iris from facial images is implemented in total.

2. Detection of eye regions in facial images Because the goal is to extract the iris pattern directly from facial images, initially a high-resolution image of the subject face is taken. In order to reduce the computational load of detecting the eye regions, the image is initially downsampled. Afterwards, the white-colored parts of the eyes are exploited to detect the eye regions within the facial image. The downsampled color image is initially converted to an HSI space as it is easier to successfully carry out a color-based segmentation in the HSI space. Thresholding is then applied to all three color components of the HSI space to get an initial segmentation result in the form of a binary image. If I H , I S , and I I are used to denote the hue, saturation, and intensity planes of

408

¨ E RT URK

the acquired image, a preliminary binary segmentation result (denoted as B1 ) is constructed in the form of   1, if t H,l ≤ I H (x, y) ≤ t H,h and t S,l ≤ I S (x, y) ≤ t S,h B1 (x, y) = , and t I,l ≤ I I (x, y) ≤ t I,h   0, otherwise (1) where (x, y) is used to denote spatial image coordinates, while t(H,S,I ),l and t(H,S,I ),h are used to respectively denote the low and high threshold values of the corresponding planes. Thresholds of each plane are preadjusted to successfully extract the white parts of the eyeball. The threshold interval of the hue component has been set to be comparatively smaller than the threshold intervals of the saturation and intensity planes, so as to enable sensitivity to white color, while still achieving robustness to illumination changes. An example result of this process is displayed in Figure 1. It is seen that the white parts of the eyes within the face region are successfully revealed. In order to dispose of speckles, morphological erosion is applied to the binary image obtained after the preliminary segmentation. For this purpose, a size 3 × 3 structuring element is utilized and large regions resistant to a single erosion operation are kept while small regions are removed. If the 3 × 3 sized structuring element is denoted as s, the erosion result can be expressed as B2 = B1  s,

(2)

where  shows morphological erosion, and B2 is used to denote the binary image obtained after erosion. Alternatively, it is possible to express the erosion process as  1, if s fits B1 B2 (x, y) = (3) 0, otherwise, which is executed for every spatial pixel location. The result of this operation is displayed in Figure 2a. It is seen that even after morphological erosion, segmented regions that are either too large (particularly in the background) or too small to make up the eye are still present. Hence, the area of segmented regions is computed at the next step, and regions that are too large or too small to make up the eye are removed. The result of this process is shown in Figure 2b. Afterwards, regions touching the image boundary are removed, and a morphological closing operation is applied to the remaining regions to combine regions that are close to each other. After close regions are combined, another area-based removal of small regions is carried out, this time with a larger threshold so that small regions not part of the eye are removed successfully. The result of these procedures is shown in Figure 3. At the next step, the following information regarding each region is collected: • The area of the region (A)

N ONINTRUSIVE I RIS I MAGE E XTRACTION

409

(a) Down-sampled (800 × 600 pixels resolution) image example.

(b) HSI domain thresholding result. Figure 1. First step in locating eye regions: HSI domain thresholding.

• • • • • • • •

The x-coordinate of the center of gravity of the region (xCoG ) The y-coordinate of the center of gravity of the region (yCoG ) The leftmost coordinate of the region (xleft ) The rightmost coordinate of the region (xright ) The upmost coordinate of the region (yup ) The downmost coordinate of the region (ydown ) The x-coordinate of the longest vertical chord (xchord ) The y-coordinate of the longest horizontal chord (ychord )

Figure 4 displays the correspondence of this information for a sample region of the white part of the eye. The eye region is then located and extracted in a four-stage approach.

410

¨ E RT URK

(a) After morphological erosion.

(b) after removing large and small regions. Figure 2. Second step in locating eye regions: Morphological erosion and area-based removal.

(1) In order to accommodate cases where both sides of the eye are located (as shown in Figure 5a), regions with close centers of gravity are combined. The information for these combined regions is reconstructed using information available about the separate regions. For example, the new area is the sum of the areas, the new leftmost coordinate is the one that is more to the left of the leftmost coordinates of the regions, the new center of gravity is the average of the centers of the gravity, etc. (2) The horizontal and vertical lengths of segmented regions are used to determine whether both sides of the eye are detected or not. If both sides of the eye are already segmented, the horizontal length should be approximately double the vertical length. Hence, if the horizontal length-to-vertical length ratio is above

N ONINTRUSIVE I RIS I MAGE E XTRACTION

411

Figure 3. Third step in locating eye regions: Removal of regions touching the image boundary and area-based thresholding after connecting close regions through morphological closing.

Figure 4. Shape information obtained about segmented regions.

a certain threshold, it is decided that both sides of the eye are detected (as in Figure 5a); otherwise, only a single side is detected (as in Figure 5b). (3) If both sides of the eye have been detected, the eye region is directly taken to be bounded between the rightmost, the leftmost, the upmost, and the downmost coordinates of the region. If a single side of the eye is detected, the x-coordinate

412

¨ E RT URK

(a) Both sides of the eye are obtained

(b) Only one side of the eye is obtained. Figure 5. Examples for detecting (a) both sides of the eye and (b) only one side of the eye.

of the longest vertical chord is used to decide which side of the eye has been detected, as the location of the longest vertical chord will always be next to the iris. Hence, if it is decided that only the left side of the eye is detected, the eye region is enlarged to the right, and viceversa. (4) Finally, the detected eye region coordinates are scaled to original size to extract the eye region from the original (high-resolution) image. With the proposed approach, the eye regions can be located using a single camera without the need to conduct a time-consuming search within the image. The proposed approach can therefore be used to detect eye regions reasonably fast, without the need of a complex system setup. Examples of eye regions extracted with the proposed approach are given in Figure 6.

N ONINTRUSIVE I RIS I MAGE E XTRACTION

413

Figure 6. Examples of eye regions located and extracted with the proposed approach.

3. Iris extraction from images of the eye region The iris is extracted from images of the eye region using only the intensity plane of the detected eye region. Initially, simple clustering-based automatic thresholding

414

¨ E RT URK

(a) Sample eye image intensity.

(b) Clustering based thresholding result. Figure 7. The effect of clustering-based automatic thresholding of eye image intensities.

is applied to the intensity planes of the eye images. The result of this operation on a sample eye image is shown in Figure 7. After clustering-based automatic thresholding, small regions are removed using morphological filling applied to the negative image. The result of this process on the binary image given in Figure 7b is shown in Figure 8. In the next step, the area of each region displayed in white in Figure 8 is computed and regions with an area below a certain threshold are removed. The result of this process separates the lower parts of the eye including the white parts, as shown in Figure 9a. Afterwards, edge detection is applied to obtain images as displayed in Figure 9b. Next, a circle is fitted to the points displayed in Figure 9b using the least

N ONINTRUSIVE I RIS I MAGE E XTRACTION

415

Figure 8. Result of removing small particles using morphological operations.

squared error criterion. In order to dispose of the effect of eyelids, the distance of all points to the initially fitted circle is computed and points with a distance above a certain threshold are removed. Then a second circle is fitted to the remainder points, and this circle is taken to identify the outer boundary of the iris using the center coordinates and radius of the fitted circle. This circle is then used to extract the iris from eye images. Sample iris patterns extracted with the proposed approach are displayed in Figure 10. Note that iris images are not always fully circular as a result of eye regions being already cut out in the form of rectangles, as explained earlier.

4. Experimental results In the experimental setup, initially a high-resolution (in our case 4048 × 3040 pixel) facial image is acquired. In order to reduce the computational load of detecting the eye regions, the image is initially downsampled to a fixed size (800 × 600 pixels in the presented results). For the HSI domain thresholding procedure, the thresholds of the hue plane are set as t H,l = 20, t H,h = 80, the thresholds of the saturation plane are set as t S,l = 0, t S,h = 210, and the thresholds of the intensity plane are set as t I,l = 50, t I,h = 250. These values are determined experimentally to give satisfactory results. Note that the threshold interval of the hue plane has been set to a comparatively smaller value compared to the threshold values of the saturation and intensity planes, so as to enable sensitivity to white color, while achieving robustness to illumination changes. Because additional size- and shape-based discrimination procedures are utilized to detect the eye region, the main task of HSI domain thresholding is to accomplish an initial segmentation. Therefore priority is given to obtain a relatively robust

Author: OK? criterion

416

¨ E RT URK

(a) After removing white regions with area below a certain threshold.

(b) Result of edge detection. Figure 9. (a) Removing small regions to separate the lower part of the eye, and (b) result of edge detection.

segmentation approach that does not overlook the eye regions, which is accomplished with relatively large threshold intervals, particularly for the saturation and intensity planes to enable, e.g., insensitivity to illumination. A total of 100 face images are captured for 25 different persons with the same experimental setup located at a fixed place. However, images are captured at different time periods, so that small illumination variations are actually present in the test set. Furthermore, subjects are allowed to stand freely in front of the camera so that images are captured at different distances, resulting in scale changes within the test set. A 100% correct extraction rate is achieved for this test set of 100 face images. Although this database is too small for conclusive evaluation, and images

N ONINTRUSIVE I RIS I MAGE E XTRACTION

417

were captured in a controlled environment and furthermore with subjects all of the same nationality (which might account for no negative samples), the results are promising. It is now commonly agreed that universal testbeds are required to assess and compare the performance of different techniques. Unfortunately, such testbeds are currently not available for many applications, which is also the case for the topic addressed in this paper. Note that the aim of the proposed techniques is to correctly identify the iris location within face images, without interest in extracting iris features for matching or detecting occluding features such as eyelids or eyelashes. Hence, eyelids and eyelashes are present in the extracted iris images, as observed in Figure 10, and additional processing, using methods similar to that proposed in [10] for instance, is required to detect such features. With nonoptimized code implemented in MATLAB, it takes 1 second to extract the eye region from the face image and 4 seconds to extract the iris from the eye region image. The main reason for the higher processing time of the iris extraction stage is the least squares circle fitting process, which is executed twice to determine the exact iris boundary. It should be possible to reduce the processing time through code optimization and implementation in a higher-level programming language.

5. Conclusions A novel approach to extract eye regions and iris patterns from facial images, taken from a distance, is proposed in this paper. The proposed approach makes use of mainly HSI domain segmentation and shape-based detection to determine the eye regions, and clustering-based automatic thresholding followed by least squares circle fitting at edges to identify iris locations within eye images. The proposed approach is simple, yet effective, as supported by the correct extraction rate. As the eye regions and the iris patterns are extracted using mostly simple and fast operations, the proposed approach can be used for nonintrusive and relatively fast extraction of iris patterns for iris recognition-based identification systems.

Acknowledgments The author thanks the research staff and students of the Electronics and Telecommunications Department of the University of Kocaeli who volunteered in the image acquisition process.

418

¨ E RT URK

Figure 10. Sample iris patterns extracted from facial images by the proposed approach (note that these images are shown at half their actual size and the image resolution is actually twice as much).

References [1] W. W. Boles, A security system based on human iris identification using wavelet transform, IEEE First Int. Conf. on Knowledge-Based Intelligent Electronic Systems, Australia, 533–541, 1997. [2] W. W. Boles and B. Boashash, A human identification technique using images of the iris and wavelet transform, IEEE Trans. Signal Process., 46, 1185–1188, 1998. [3] T. A. Camus and R. Wildes, Reliable and fast eye finding in close-up images, 16th International Conference on Pattern Recognition, 389–394, 2002. [4] B. Cao, S. Shan, W. Gao, and D. Zhao, Localizing the iris center by region growing search, IEEE International Conference on Multimedia and Expo. ICME ’02, 129–132, 2002. [5] J. G. Daugman, High confidence visual recognition of persons by a test of statistical independence, IEEE Trans. Pattern Anal. Machine Intelligence, 15, 1148–1161, 1993. [6] H. M. El-Bakry, Fast iris detection for personal identification using modular neural networks, IEEE International Symposium on Circuits and Systems, ISCAS 2001, 581–584, 2001. [7] Y.-P. Huang, S.-W. Luo, and E.-Y. Chen, An efficient iris recognition system, First Int. Conf. on Machine Learning and Cybernetics, Beijing, 450–454, 2002.

N ONINTRUSIVE I RIS I MAGE E XTRACTION

419

[8] T. Kawaguchi, D. Hidaka, and M. Rizon, Detection of eyes from human faces by Hough transform and separability filter, IEEE International Conference on Image Processing, 49–52, 2000. [9] W. K. Kong and D. Zhang, Accurate iris segmentation based on novel reflection and eyelash detection method, Int. Symp. on Intelligent Multimedia, Video, and Speech Processing, Hong Kong, 263–266, 2001. [10] Y. Liu, S. Yuan, X. Zhu, and Q. Cui, A practical iris acquisition system and a fast edges locating algorithm in iris recognition, Instrumentation and Measurement Technology Conference, 166– 168, 2003. [11] L. Ma, T. Tan, Y. Wang, and D. Zhang, Personal identification based on iris texture analysis, IEEE Trans. Pattern Anal. Machine Intelligence, 2, 1519–1533, 2003. [12] L. Ma, Y. Wang, and T. Tan, Iris recognition using circular symmetric filters, 16th International Conference on Pattern Recognition, 414–417, 2002. [13] M. Negin, T. A. Chmielewski, Jr., M. Salganicoff, T. A. Camus, U.M.C. von Selen, P. L. Venetianer, and G. G. Zhang, An iris biometric system for public and personal use, IEEE Computer J., 70–75, 2000. [14] R. Sanchez-Reillo, C. Sanchez-Avilla, and J. A. Martin-Pereda, Minimal template size for iris recognition, BMES/EMBS Conf. Serving Humanity, Advancing Technology, Atlanta, CA, USA, 972, 1999. [15] K. Toennies, F. Behrens, and M. Aurnhammer, Feasibility of hough-transform-based iris localisation for real-time applications, 16th International Conference on Pattern Recognition, 1053–1056, 2002. [16] R. P. Wildes, J. C. Asmuth, G. L. Green, S. C. Hsu, R. J. Kolczynski, J. R. Matey, and S. E. McBride, A system for automated iris recognition, Second IEEE Workshop on Application of Computer Vision, 121–128, 1994. [17] G. O. Williams, Iris recognition technology, IEEE AES Systems Magazine, 23–29, 1997. [18] L. Yu, K.-Q. Wang, C.-F. Wang, and D. Zhang, Iris verification based on fractional Fourier transform, First Int. Conf. on Machine Learning and Cybernetics, Beijing, 1470–1473, 2002. [19] Y. Zhy, T. Tan, and Y. Wang, biometric personal identification based on iris patterns, 15th Int. Conf. on Pattern Recognition, 801–804, 2000.