1192

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

Iris Data Indexing Method Using Gabor Energy Features Somnath Dey, Student Member, IEEE, and Debasis Samanta, Member, IEEE

Abstract—Biometric features are extracted from a complex pattern and stored as high dimensional data. These data do not follow traditional sorting order like numerical and alphabetical data. Hence, a linear search method makes the identification process extremely slow as well as increases the false acceptance rate beyond an acceptable range. To address this problem, we propose an efficient indexing mechanism to retrieve iris biometric templates using Gabor energy features. The Gabor energy features are calculated from the preprocessed iris texture in different scales and orientations to generate a 12-dimensional index key for an iris template. An index space is created based on the values of index keys of all individuals. A candidate set is retrieved from the index space based on the values of query index key. Next, we rank the retrieved candidates according to their occurrences. If the identity of the query template is matched, then it is a hit, otherwise a miss. We have experimented our approach with Bath, CASIA-V3-Interval, CASIA-V4-Thousand, MMU2, and WVU iris databases. Our proposed approach gives 11.3%, 14.5%, 16.3%, 13.5%, and 10.3% penetration rates and 98.2%, 91.1%, 90.7%, 85.2%, and 96% hit rates for Bath, CASIA-V3-Interval, CASIA-V4-Thousand, MMU2, and WVU iris database, respectively, when we consider the retrieving templates up to the fifth rank. Experiments substantiate that our approach is capable of retrieving biometric data with a higher hit rate and lower penetration rate compared to the existing approaches. Application of Gabor energy features to index iris data proves to be effective for fast and accurate retrieval. With our proposed approach, it is possible to retrieve a set of iris templates similar to the query template in the order of milliseconds and is independent of sizes of databases. Index Terms—Gabor energy calculation, image retrieval, indexing biometric data, iris biometric, personal identification.

I. INTRODUCTION

O

F LATE, biometric authentication systems are being deployed in many applications [1] such as PDAs, smart cards [2]–[5], network access [6], [7], biometric passports [8], security screening [1], [9], [10], database access, user authentication [1], etc. Demands are increasing to deal with large-scale databases in these applications [1], [9], [10]. For example, the Unique Identification Authority of India (UIDAI) has planned to register 600 million users in India in the next few years [11], [12] where the number of accesses per day (in different public and private domains) are expected to be around 1 to 5 million. Manuscript received January 06, 2012; revised April 07, 2012; accepted April 11, 2012. Date of publication April 25, 2012; date of current version July 09, 2012. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Jaihie Kim. The authors are with the School of Information Technology, IIT Kharagpur, 721302, West Bengal, India (e-mail: [email protected]; [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TIFS.2012.2196515

Two problems may arise in a large-scale biometric database: 1) searching the database to retrieve an identity can be slow because, in the worst case, the input biometric data has to be compared (matched) against every biometric data in the database, and 2) the false acceptance error grows with the sizes of databases [13]. As a consequence, the response time, search, and retrieval efficiency are affected in addition to the accuracy of the system. To alleviate these limitations, an indexing technique has been proposed to retrieve the data from large biometric databases [13]–[22]. We may note that in traditional databases, records are indexed in an alphabetical or numeric order with respect to primary key field(s). Such a technique is not applicable to index biometric databases because the biometric data do not have a natural sorting order. As a way out, indexing schemes in biometric databases would assign an index value to each biometric template. However, with this method, sometimes the indices of two biometric templates related to the same identity may differ due to noise involved in the process of data acquisition and thus poses a concern on accuracy. This work addresses this problem and investigates for a fast and accurate mechanism to index iris databases. As far as iris database indexing is concerned, it is a challenging task because of its high dimensional features set. Several texture feature extraction methods are known for iris-based authentication in verification mode [23]–[29]. Nevertheless, retrieval of biometric data with a large number features stored in a huge pool of memory are scarcely reported. Mukherjee and Ross [14] propose two methods for iris data indexing. In their first method, they calculate IrisCode features [23] using Gabor wavelet. IrisCode features, according to their method, are as large as 2048 dimensions. To reduce the dimensionality of features, they proposed row/column/block averaging and principle component analysis (PCA). They propose another technique called signed pixel level difference histogram (SPLDH) analysis [14]. In this technique, they divide an iris texture into a number of blocks. For each block they calculate the histogram of signed differences pixel intensities (of similar positioned pixels in the block and adjacent blocks). Like the IrisCode method, this method also deals with features of around 256 dimensions. Further, determining the dimensionality of the histogram remains an issue in this technique. The scale invariant feature transform (SIFT) method [16] is also known to extract texture features for iris data indexing. This method finds a number of key points. Then high dimensional (124 dimensions) SIFT features are extracted for each key point. The key points are indexed using geometric hashing. This indexing method also deals with a large number of key points (approximately 100 key points) with high dimensional feature vector.

1556-6013/$31.00 © 2012 IEEE

DEY AND SAMANTA: IRIS DATA INDEXING METHOD USING GABOR ENERGY FEATURES

Local binary pattern (LBP) [30] is another method in the field of iris biometric identification. In his thesis, Mukherjee reports the experiment on block-based LBP in iris indexing. They apply blockwise LBP to extract iris features and then -means clustering is followed to index iris data. The limitation of the IrisCode (row/column/block averaging, PCA), SPLDH, and SIFT methods is that all of them deal with high dimensional features which make the methods computationally expensive and suspect to statistical unreliability. It is reported that in the IrisCode-based method, an 80% hit rate is achieved at a 21%, 8%, and 17% penetration rate for row/ column averaging, block-based averaging, and PCA, respectively [14]. In the SPLDH-based method, an 84% hit rate is achieved at a 30% penetration rate [14]. Further, with the SIFT method, a 61% hit rate is achieved at a 37% penetration rate [16]. It is observed that to achieve an 80% hit rate, the average penetration rate is around 70% according to the LBP method [30]. Moreover, LBP does not represent the iris texture well to be considered as a representative feature for indexing. We may note that, LBP represents only the local contrast and in fact does not capture any interblock relationship. Another limitation is that the most of the indexing techniques are based on partitioning of biometric databases [14], [30]. However, partitioning methods do not equally divide the databases. Since the sizes of the partitions are not uniform, searching a larger partition does not give the same result as searching a smaller partition. Hence, techniques are usually neither efficient nor accurate enough to be applicable in practical applications. In essence, it is desirable to have an indexing technique which not only reduces the search space but also guarantees the accuracy. Some techniques [31]–[38] are also available to deal with image retrieval from large image databases. But these techniques cannot be applied to iris images because the dimensionality of the feature vector is large and the feature representation of biometric data is different. In this paper, we propose an indexing technique for an iris-based biometric identification system. In our proposed approach, we find the iris part from an eye image. The texture features are extracted from the iris image using a Gabor filter in different orientations and different scales. We calculate the Gabor energy features from these texture features and then create the index key values. Gabor features have optimal localization properties in both spatial and frequency domains and are invariance to illumination, rotation, scale, and translation of image. The multichannel Gabor function has been recognized to be very useful in the field of image retrieval. The dimensionality of the multichannel Gabor features is reduced by calculating the Gabor energy. This motivates us to use Gabor energy features to index iris data. The reduced low dimensional feature vectors make the proposed approach computationally fast. We also propose a storing and searching mechanism for the iris biometric database based on the indices. Our indexing method finds a list of similar templates on the order of milliseconds irrespective of the sizes of databases. The experimental results show faster access with better accuracy in terms of higher hit rate and lower penetration rate. The rest of the paper is organized as follows. A basic concept of Gabor filter is discussed in Section II. In Section III, we de-

1193

scribe our proposed methodology. The experimental results are presented in Section IV. Finally, Section V concludes the paper. II. PRELIMINARIES We apply a Gabor filter in our work. For better understanding of our work, we briefly discuss the concept of the Gabor filter in this section. Gabor transform theory was first proposed by Gabor in 1946 [39]. Daugman [40] proposed two-dimensional (2-D) Gabor transform theory in 1985. It has been observed that the 2-D Gabor filter is an effective method for time-frequency analysis [40], [41]. In the spatial domain, a 2-D Gabor filter consists of a sinusoidal wave modulated by a Gaussian envelope. It performs a localized and oriented frequency analysis of a 2-D signal. A 2-D Gabor function and its Fourier transform [34], [35] can be expressed as

(1) (2) and . In (1), and are where the standard deviations of the Gaussian envelope along and directions and determine the filter bandwidth. Here, and are the spatial domain and frequency domain coordinates, respectively. is the center frequency of the filter in frequency domain. A multiresolution Gabor filter (also called multichannel Gabor wavelet) is a set of filter banks with different scales (frequencies) and different orientations. The Gabor function forms a complete but nonorthogonal basis set. Expanding a signal using this basis provides a localized frequency description, thus capturing local features or energy of the signal. Texture features can then be extracted from this group of energy distributions. The Gabor wavelet can be represented with a mother wavelet and derived the appropriate Gabor functions by different rotations and scales. Let be the mother wavelet of the Gabor filter family. The set of Gabor filter bank can be generated from as expressed in (3) (3) where is a constant such that variables, and

,

and

are two integer

(4) , , and In (4), . The parameters and are the total number of scales and orientations in the multiresolution decomposition, respectively. The values of and are to be chosen to ensure that the halfpeak magnitude support of the filter responses in the frequency

1194

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

spectrum touch each other. The filter parameters , can be computed as follows:

, and

Fig. 1. Overview of our proposed approach.

and (5) denote the lower and upper center frequencies where and of interest, respectively. The Gabor feature space consists of responses calculated with a multiresolution Gabor filter at several different scales (frequencies) and orientations. The response of a Gabor filter to an image is obtained by a 2-D convolution. Convolution of an image with the kernel gives a response that is proportional to how well the local features in the image match the kernel. Let denote an image and denote the response of a Gabor filter at the th scale and th orientation to an image point . The Gabor filter response to an image is defined as follows: (6) The Gabor filtered image has both real and imaginary components as the response of the Gabor filter is complex. The magnitude of the Gabor filtered image is calculated using (7) and are the real and where imaginary parts of the Gabor filtered image, respectively. In this work, we use multiresolution Gabor filter to extract features from iris images. III. PROPOSED METHODOLOGY In this section, our proposed methodology for iris biometric database indexing is discussed. An overview of our proposed approach is shown in Fig. 1. The different tasks involved in our approach are discussed in the following. A. Preprocessing We propose iris localization and normalization as the preprocessing tasks for a captured eye image. First, in localization, the iris part is isolated from an eye image. To localize the iris part, we preprocess the eye image by applying downscaling and color level transform. The downscaling of the eye image is done to reduce the search area for pupil and iris boundaries. The original eye image is with the resolution 320 280 and the eye image is downscaled to the resolution 160 140. Fig. 2(a) shows a sample eye image. The color level transform is applied to minimize the influence of irrelevant edges. In other words, the color level transform increases the intensity differences between the pupil and iris parts, and iris and sclera parts which in turn helps

Fig. 2. Preprocessing result of a sample iris image [29], [42]; (a) original eye image; (b) pupil boundary detected eye image; (c) iris boundary detected eye image; (d) localized iris image; (e) normalized iris image; (f) enhanced normalized iris image.

to detect the pupil and iris boundaries efficiently and accurately [42], [43]. The next step is to find the pupil boundary. To detect the pupil boundary, we convert the color level transformed image to the binary image and find all connected components in the binary image. Then we remove the small irrelevant components which may occur due to eyelashes, eyelids, and noise. The pupil component is selected by counting the number of pixels within the calculated average radius. The pupil boundary is found from the pupil component by edge detecting and edge connecting. The pupil centroid is determined by calculating the centroid of all pixels within the pupil boundary. Fig. 2(b) shows a pupil detected image. Iris boundary detection is the next step in the iris localization. To detect the iris boundary, the eye image is divided into left and right images at the pupil centroid. Color level transform is applied on both left and right images to enhance the contrast between iris and sclera boundaries. The image preprocessing technique, namely dilation [44], is followed to color level transformed image to reduce the noise effect in edge detection. The dilated image is then thresholded to create a binary image and to detect the vertical edges. These edges mainly occurred due to iris boundary. Irrelevant edges which occur due to noise and eyelashes are removed after the edge detection. Iris boundary and eyelid boundary are detected by checking the pixel connectivity and drawing the small lines in particular directions. Fig. 2(c) shows detected iris boundary in the eye image. Finally, resizing of pupil and iris boundary information is done. The detected iris part is shown in Fig. 2(d). Details of the pupil and iris boundary detection have been reported in [42] and [43]. After completion of iris localization, the iris part is wrapped into a fixed size rectangular block to make the iris sample scale and translation invariant. This process is called iris normalization. A detail about iris image normalization technique can be found in [29]. The wrapping process is done by transforming the iris texture from Cartesian to polar coordinate. The Daugman’s homogeneous rubber sheet model [23] is applied to normalize the iris texture. The normalized iris image is shown in Fig. 2(e).

DEY AND SAMANTA: IRIS DATA INDEXING METHOD USING GABOR ENERGY FEATURES

1195

Then the normalized iris is enhanced to make the iris texture illumination invariant by applying techniques proposed in Ma et al. [24]. This enhanced image is used in the feature extraction. The enhanced normalized iris image is shown in Fig. 2(f). B. Feature Extraction We extract Gabor energy features in different scales and orientations from a normalized iris image as stated below. We know that the response of a Gabor filter to an image is obtained by a 2-D convolution operation. Let be the normalized iris image and denotes the response of Gabor filter in the th scale and th orientation to an image at point on the image plane. The Gabor filtered image can then be obtained using (6) as stated in Section II. We apply the multiresolution Gabor filter to extract the iris texture features. For this purpose, we apply the Gabor filter in each scale and orientation, and compute the response at each position of the image. These responses are called Gabor coefficient values which are complex. We calculate the magnitude values of the responded image at each pixel. The Gabor energy is calculated by summing up the square values of the magnitude of Gabor responses at each pixel as in

For a given iris image, we create its index key as mentioned earlier. To store any index key for any sample we propose an index organization as shown in Fig. 3. The organization consists of number of tables. Each table is corresponding to each feature of the index key. The length of a table for a feature depends on the minimum and maximum values for all samples of all individuals. Thus, the length of the table for any th feature is calculated using

(8)

(11)

It is observed that the range of values of Gabor energy features is very high. It is also examined that intrafeature range distribution is higher for lower values of and . To make the intrafeature range distribution similar, we have studied that a nonlinear function such as can serve the purpose, that is, it effectively maps the high range of Gabor energy features’ values to a lower range of values. Thus, for a given iris image we represent the Gabor energy features in the form of a matrix as follows:

.. .

.. .

Fig. 3. Index key organization.

D. Database Creation

and are the maximum and minimum values where for all samples of all individuals of the th feature. In (11), varies from 1 to . Suppose, there are number of individuals and for each individual there are number of samples. Let denotes the th feature of the th sample of the th individual. The and are then calculated using

(12)

.. .

(9) The rows and columns denote the number of scales and number of orientations of the multiresolution Gabor filter. We may note that the feature vector in our approach consists of number of features. C. Index Key Generation The extracted Gabor energy features in different scales and orientations are used to generate a key for iris database indexing. In our approach, a key represents a feature vector and a feature vector consists of number of features’ values. We denote the th value of an index key as as follows: (10) represents the logarithm value of the Gabor enwhere ergy feature in the th scale and th orientation. In (10), and and . In other words, the th feature of an index key represents the logarithmic value of a Gabor energy feature in a particular scale and orientation. Note that the length of the index key is . Let us denote this length as .

The index value of the table corresponding to the th feature is started with 0 and ends with . The value of is calculated using (13) Each entry of a table contains two lists (see Fig. 3). The list contains a set of values which we call the median values of all samples for an individual of a feature representing the table and the list contains a set of image names. An image information corresponding to a sample is uniquely identified by this image name. 1) Memory Usage: Our proposed approach considers a number of tables to index iris data. We analyze the amount of memory overhead for the proposed indexing mechanism. In our database, each table stores a sample one time and each table has entries for all samples. Each entry requires 4 bytes. Hence, the memory requirement to store all index keys is calculated using (14) denotes the size of a table, is the number In (14), of features used for indexing, is the number of samples in the database, and is the memory required to store a sample.

1196

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

E. Storing Iris Biometric Index Key To store the index keys in the database, we sort the feature values for a feature corresponding to all given samples of an individual. Then we find the center value of the feature values for that individual. Note that median or mean value can give us the center value. The median value gives the central position which minimizes the average of the absolute deviations. On the other hand, the mean value gives the center position biasing toward the extreme value. For example, if some feature value is very low or very high for a particular sample, then the mean value is close to the feature value of that sample. Hence, we use the median value in our approach. For any th feature, we compute the median value given number of samples for an individual say using (15) Note that if there are number of individuals and each individual has number of samples, then we store number of median values and image names in each table. Let us denote the unique identifier for the th sample of the th individual as and the th feature value as . We store the corresponding to the at th position median value and in the th table. Given the value of , the value of index , in the th table is calculated using Fig. 4. (a) Average table size and (b) memory requirements to enroll iris data for different iris image databases.

Fig. 4 shows the average table size and memory requirement to store the different numbers of samples for different databases. From Fig. 4(a), we see that the table size does not vary linearly when the number of enrolled samples increases. Further, Fig. 4(b) shows that the memory requirement to store samples is increased linearly. This result substantiates that the total memory is not increasing nonpolynomially. We have calculated the memory cost for 1 000 000 and 1 000 000 000 individuals according to (14). The size of a table depends on the range of values of Gabor energy features. From Fig. 4(b), we observe that the size of a table is approximately 1.2 times the number of samples (according to our experiment with 1000 enrollments). We may note that beyond 1000 enrollments the size of the table does not increase significantly. Considering this, for large enrollments, we may reasonably assume the size of the table as 1.5 times the number of individuals. Further, we assume that 4 bytes is required to store an individual. With this consideration, we estimate the memory requirement for 1 000 000 and 1 000 000 000 enrollments are as follows:

Hence, to store the above-mentioned data proposed approach requires approximately 115-MB and 112-GB storage, respectively.

(16) If we want to enroll additional samples of a subject which is already enrolled in the database, we recalculate the median value of all enrolled samples with the additional samples for that subject. Then, we enroll the all samples of the subject into the database. If we want to enroll a new subject, first we calculate the median value of each feature from all samples of the new subject. Then we check the minimum and maximum values of each feature whether the values are within the range of the table or not. If feature values lie within the range of the table then we enroll the identity and the median value of the new subject. If the values do not fall within the range, then we increase the size of the table with the new minimum and maximum values and reorganize the enrolled samples. Finally, we enroll all samples of the new subject. 1) Illustration: We illustrate our approach to store iris biometric data and index them with an example. Suppose, we are to store the data of ten individuals and each individual includes five sample data. Thus, the total number of samples to be stored and indexed is 50. Further, we consider three scales and four orientations for Gabor feature extraction. Hence, the length of the index key is 12. That is, we need 12 tables in the database to store the 12 features for each index key. The index keys for 50 samples are created as discussed in Section III-C. Fig. 5 shows the index keys of five samples and all data pertaining to say, the sixth individual for an instance. Let us assume that the minimum and maximum values of the first feature analyzing all 50 samples are found to be 100 and 350, respectively. Thus, the length of the first table in the database is 251 . Similarly, the length of the other tables in the database can be calculated from maximum and minimum values of features. Now, we want to enroll the first sample of the sixth individual into the database. Among all the index keys of the sixth individual, the first sample is highlighted as shown in Fig. 5. Now, we are to

DEY AND SAMANTA: IRIS DATA INDEXING METHOD USING GABOR ENERGY FEATURES

1197

Algorithm 2 Enrollment all samples for an individual into the database

Fig. 5. Index keys of the sixth individual with five samples.

find the median value for each feature of the sixth individual. The values of the first feature of the sixth individual are shown as a rectangular block in Fig. 5. From Fig. 5, we can see that the median value of the first feature of the sixth individual is 318. The first feature value of the first sample of the sixth individual is 316. We store the median value (318) of the sixth individual in the and the identity of the first sample of the sixth individual in the at the 216th (316–100) location of the first table of the database, respectively. We store the other features of the first sample of the sixth individual likewise. We can enroll the rest of the samples in the same way. Algorithm 1 formally states the techniques of creating tables in the database. The enrollment of the index key of a sample for an individual is summarized in Algorithm 2. We use the following notations in our algorithms.

Notations used in our algorithms Number of individuals. Number of samples of each individual. Number of features in an index key. is the th feature of th sample of th individual. is the th iris image of th individual. is the list of median values. is the list of pointers to iris image features. is minimum value for th feature. is maximum value for th feature. is the table to store index key of the th feature. is the set for retrieved iris template corresponding to th feature. Algorithm 1 Creating tables for indexing 1: for to do 2: for all do 3: for all do 4: //Finding Maximum of all samples of all individuals 5: //Finding Minimum of all samples of all individuals 6: end for 7: end for 8: //Calculate the length of the table for each feature 9: Create of length //Create the table for each feature 10: end for

1: Let the individual be 2: for to do 3: 4: end for 5: //Store feature vector into database 6: for all do 7: for to do 8: //Calculate the index for feature value 9: Add to 10: Add to 11: end for 12: end for F. Retrieving the Best Match(es) Once all individuals are enrolled into the database, we can use it to retrieve data for a match. As a task of retrieving, we are to find the iris templates from the database which are the most similar to the query iris template. To do this, we first generate the index key of length from the query iris as discussed in Section III-C. Note that we have only a set of feature values for the given query iris image. There is no median value as we do not know the median value a priori. At the time of retrieval, we retrieve a set of median values and sample names corresponding to each feature value. For the th feature (with feature value say ) of the query index key, we retrieve a particular position of the th table in the database based on the value of . The position of the th table is decided by (17) at the We retrieve a set of iris image names location of the th table. Let this set be . We also find the minimum and maximum median values from the list at location of the th table. Next, we retrieve the sets corresponding to the values to from and add these names to . Here, is a threshold value which would be decided empirically. In this way, we retrieve the candidate iris images for the th feature and store them in a temporary list . Similarly, we retrieve the iris image names for other features and store them in respective temporary lists. In our next step, we find the most common iris images among all sets ’s and compute ranking for each name. First, we merge all ’s to ; that is, . To count the number of occurrences of each individual, we maintain a counter for each individual which is in the set . We increase the counter corresponding to an individual when it occurs in the set . Then we sort all the counters in descending order. We assign rank one to the individual corresponding to the highest value of the counter, rank two to the next highest values of the counter, and so on. It means that the maximum number of occurrences is assigned as rank one and other ranks are given in descending order of number of occurrences. The highest rank indicates that the retrieved iris image is most similar to the query image.

1198

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

Therefore, the range of locations is 214 (315 100 1) to 219 (318 100 1). We add all iris image names in at locations 214 to 219 into . The set is shown in Fig. 6(c). Similarly, we create sets for the rest of the features. We avoid the showing of the iris image names’ position in the other tables due to clarity of the figure and the limitation of space in the paper. For continuity of our illustration, we assume the other 11 sets of retrieved iris image names, as shown in Fig. 6(d). We merge all such sets to . Finally, we calculate the number of occurrences of particular in and calculate the rank. To do this, we maintain a counter for each individual which is in the set . We increment the counter corresponding to an individual when it occurs in the set . Fig. 6(e) shows the counter values for each image name. Fig. 6(f) shows the individuals with corresponding rank and the sixth individual being the rank 1 will be retrieved as the best match. The steps for retrieving iris template from a database are summarized in Algorithm 3. Algorithm 3 Retrieving iris template from database 1: Let the query template be 2: for to do 3: Let the th feature value in be 4: //Calculate the table location 5: 6: 7: 8: Fig. 6. Retrieving a match for a query iris image from a database; (a) index table of the database for first feature; (b) index key vector for query iris template; for first feature; (d) sets for other 11 features; (e) counter values for (c) set each individual present in sets; (f) individuals with their ranks.

1) Illustration: We illustrate how we can retrieve a match for a given query iris image. We consider that the index keys of ten individuals and five samples for each individual are stored in our database. Thus, there are 50 index keys in the said database. Let us assume the minimum and maximum values of the first feature (with respect to these 50 index keys) are 100 and 350, respectively. The length of the first table in the database is 251 . Fig. 6(a) shows the first table for the first feature. Now, we consider a query iris template . We generate the index key from the query template. In this example, let us consider three scales and four orientations for Gabor feature extraction as in the storing time. Hence, the length of the index key is 12. The index key generated from the query iris template of length 12 is shown in Fig. 6(b). Here, the first feature value of the query iris template is 317 [see Fig. 6(b)]. Knowing this, we retrieve all median values in the at location 217 of the first table. We see that the median values in the at the 217th location of ’s table are 315 and 318 [see Fig. 6(a)]. The minimum and maximum median values in the are 315 and 318, respectively. We also retrieve all iris image names in the at location 217 (317 100) to . Next, we add the iris image names in the from a range of locations based on minimum and maximum median values and preassigned threshold value . Let the value of be chosen as 1.

9: 10: 11: 12: 13: 14: 15:

for

to

do

//Add iris image names to end for //Merge all set to end for Define counter for each individual in set . Initialize all to zero for Each do //Calculate the number of occurrence of each individuals

16: 17: end for 18: Sort all in descending order 19: Assign rank based on the value of IV. EXPERIMENTAL RESULTS To study the efficacy of the proposed iris data indexing approach, we have conducted a number of experiments. In this section, we present the experiments carried out and the experimental results observed. A. Experimental Setup In our experiment, we have used four different iris image databases: Bath University (Bath) [45], CASIA-IrisV3-Interval (CASIAV3I) [46], CASIA-IrisV4-Thousand (CASIAV4T) [47], Multimedia University (MMU2) [48], and West Virginia University (WVU) [49]. The Bath database contains 1000 eye images of 25 persons. Each person has 20 images of left eyes and 20 images of right eyes. The CASIAV3I database contains 2639

DEY AND SAMANTA: IRIS DATA INDEXING METHOD USING GABOR ENERGY FEATURES

TABLE I CHARACTERISTICS OF DATABASES USED IN OUR EXPERIMENTS

eye images of 249 persons. The eye images in the CASIAV3I database are captured from 395 eyes. The CASIAV4T database contains 200 000 eye images of 1000 persons and each person has 10 left and 10 right eye images. The MMU2 database contains 995 eye images from 100 persons. Five images are captured from each eye of a person. There are five left eye images which are excluded from the database due to cataract disease. The WVU database contains 3099 eye images from 244 persons. Left and right eye images are captured from 241 and 236 persons, respectively. In the WVU database, 1–20 samples are captured for each eye. It may be noted that the iris data of left and right eyes of a person are different [50]; we therefore consider 50 unique individuals in the Bath database, 395 unique individuals in the CASIAV3I database, 2000 unique individuals in the CASIAV4T database, 199 unique individuals in the MMU2 database, and 477 unique individuals in the WVU database in our experiment. A summary of all the databases is given in Table I. We have done our experiments with an Intel Core2Duo processor (2.00 GHz) and 2.0-GB memory. We use GCC 4.3 compiler to develop our program. B. Experimental Results With the above-mentioned experimental setup, we analyze the enrollment time, searching time, and efficiency of our proposed approach. Our observations are summarized in the following. 1) Enrollment Time: We have done the run time analysis for enrollment of samples of a new subject with big-O notation. Let be the total number of samples (from number of individuals) enrolled into the database. Let there be number of samples for an individual. Now, we want to enroll another number of samples of th individual. We calculate the index keys of these new samples using our proposed index key generation method (Section III-C). Each index key contains 12 features. We calculate the median value of each feature which takes computation. Then we check the minimum and maximum values of each feature whether the values are within the range of the table or not. If feature values lie within the range of the table, then we enroll each sample into the table. We enroll individual’s identity and the median value of the th feature of a sample into the th table. The position in the table is calculated based on the feature value with time complexity (see (16), Section III-E). If the values do not fall within the range, then we increase the table size with the new minimum and maximum values and reorganize the enrolled samples. The time complexity of this process is . 2) Searching Time: We analyze the run-time complexity with big-O notation for gallery match corresponding to a query sample. Let be the total number of samples enrolled in the database (from number of individuals) and

1199

an individual has number of samples. With reference to Algorithm 3, given a query template, we retrieve the identities of individuals and median values stored at position from the th table . is calculated using (17). The time complexity of this process is . Next, we retrieve the sets corresponding to the values to from and add these names to for rank calculation. This process can be accomplished in time. Let be the average size of the list. The rank calculation process can be done in time. In the worst case, when all samples are stored in one position, then is equal to , which is very unlikely to occur. We analyze the search efficiency by measuring the average time taken to retrieve iris templates from the database for a given query iris. Let be the average time to perform addition, subtraction, and assignment operations. Our indexing approach requires six comparisons to retrieve candidates corresponding to one feature and a candidate set of size is retrieved using 12 features (see Algorithm 3). Therefore, the time taken to retrieve a candidate set of size is . Let be the time to compare query iris template with one stored iris template for matching. Hence, the search time using our proposed indexing approach is . On the other hand, linear search requires time. Thus, our indexing approach takes less time than the linear search approach because . 3) Efficiency: To analyze the efficiency of the proposed approach, we divide all the samples into two sets: Gallery Set and Probe Set. The Gallery Set contains 80% samples of the all identities and the other 20% is considered for the Probe Set. The samples for the Gallery Set are chosen randomly. We enroll all samples of the Gallery Set into the database. Next, we find the best match for each sample in the Probe Set. In our experiment, we measure the performance of our approach in terms of two metrics: Hit rate (HR) and Penetration rate (PR). HR is defined as the ratio of correctly probed samples and the total number of probes. PR is defined as the fraction of user samples retrieved from the database upon presentation of a query template. Let be the number of correctly probed samples among number of probes and be the number of samples retrieved from a database of size for an th probe. Then HR and PR are defined in (18) and (19) as follows:

(18) (19) We analyze the best match for each sample in the Probe Set at different ranks. We calculate the HR and PR for different ranks and summarize the results in Table II. We consider different ranks between 2 and 20 of the iris images as each individual has a maximum of 20 samples. In Table II, we see that the HRs for CASIAV3I, CASIAV4T, MMU2, and WVU are considerably decreased at higher ranks. The HRs decrease because the number of samples for an individual are less for CASIAV3I, CASIAV4T, MMU2, and WVU databases (see Table I). We may note that the average PR is 11.6%, 15.3%, 16.9%, 14.3%, and

1200

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

TABLE II HR AND PR FOR DIFFERENT IRIS IMAGE DATABASES

Fig. 8. ROC curve for CASIA-V3-Interval database in semilog scale.

Fig. 7. CMC curves for different databases.

10.8% for Bath, CASIAV3I, CASIAV4T, MMU2, and WVU iris databases, respectively. We also substantiate our result in terms of cumulative match score which gives the probability of at least one correct identity present within a top rank. How cumulative match score varies with rank is shown in Fig. 7 as cumulative match characteristics (CMC) curve. It is obvious that cumulative match score increases with the increase of rank. From Fig. 7, we see that 99.4%, 98.8%, 98.7%, 98.9%, and 98.8% HRs are possible within the top 60 ranks for Bath, CASIAV3I, CASIAV4T, MMU2, and WVU databases, respectively. In the existing literature on iris indexing, the best reported result [16] so far is 98.56% HR (cumulative match score) within the top 100 rank for the CASIAV3I iris database. In other words, if we consider the top 100 rank in our approach, then the HR (cumulative match score) will eventually increase. Further, we analyze the performance of the proposed method with respect to False Positive Identification Rate (FPIR) and False Negative Identification Rate (FNIR) on the CASIAV3I database. To do this first we calculate False Match Rate (FMR) and the False NonMatch Rate (FNMR) as follows. We match each query template of the probe set with each template in the gallery set using Daugman’s [23] iris recognition method. We choose 3000 genuine pairs and 1 242 608 imposter pairs from the gallery and probe of the CASIAV3I database. We calculate the genuine score and imposter score using Daugman’s method [23] for each genuine and imposter pairs, respectively. The trade-off between FMR and FNMR is shown by the receiver operating characteristic (ROC) curve in Fig. 8. The equal error rate (ERR) of the system is 0.19% on the CASIAV3I database. Finally, we calculate FNIR and FPIR for an identification system [51] without indexing and with indexing using (20) and (21), respectively, (20)

Fig. 9. FPIR versus FNIR curve with indexing and without indexing for CASIA-V3-Interval database.

(21) In (20), is the number of templates in the database and in (21), is the PR and PRR is the indexing error. We calculate PRR using (22) where CMS is the cumulative match score. The trade-off between FPIR and FNIR for the identification system without indexing and with indexing is shown in Fig. 9. From our experimental result, it may be interpreted that with 0.0019% FMR and 3.83% FNMR we can achieve 0.72% FPIR and 3.95% FNIR with indexing and 3.93% FPIR and 3.83% FNIR without indexing. From Fig. 9, we can observe that using our proposed indexing approach we can achieve low FPIR for an FNMR. We compare our work with some of the existing indexing techniques [14] and [16]. To compare our approach, we consider the top five matches for all databases. Note that existing approaches do not follow rank-based evaluation. Hence, we compare the result of the existing work with the same at the fifth rank of our algorithm with respect to HR and PR. We consider the fifth rank in our approach because in all databases, a minimum five samples of a subject are present. Table III shows the comparison results. From Table III, we see that our approach gives better HR than the approaches proposed in [14] and [16].

DEY AND SAMANTA: IRIS DATA INDEXING METHOD USING GABOR ENERGY FEATURES

TABLE III COMPARISON OF HR AND PR WITH EXISTING WORK

It may be noted that the PR according to our approach is comparable to the existing indexing techniques. We also analyze the existing algorithms to compare the retrieval times. In the indexing method, a set of templates is retrieved for a given query template and then searching or matching is performed on the retrieved data. The efficiency of the indexing system can be measured in terms of the retrieval time of the system. We have analyzed the retrieval times of existing indexing techniques which is stated as follows. The IrisCode-based method [14], [30] stores iris data into a number of clusters using -means clustering. For a given query sample, the IrisCode method determines the target cluster for the closest match. This indicates that the query template is needed to compare with all clustered centers to find a matched cluster. This technique requires a minimum number of comparisons to find the cluster where is the number of clusters. Typical values of for 2000 samples and may increase when the number of enrolled samples increases. Hence, the IrisCode-based technique requires number of computations to retrieve similar iris templates. It also may be noted that the number of varies with the sizes of databases [14], [30]. The SPLDH-based approach [14] uses a tree data structure to store iris data. The minimum search time for the tree data structure to find a match with query template is which is the height of the tree with number of enrollments. This is so because the height of the tree depends on the number of samples stored in the tree. The SIFT-based technique [16] uses geometric hashing for indexing. In SIFT-based indexing, a minimum number of comparisons are required to hash all key points where is the number of key points in the query sample and the value of is approximately 100. This time is independent of the sizes of databases. On the other hand, in our approach, we retrieve the similar templates corresponding to 12 index key values of query template. This requires 12 comparisons in time. Further, the retrieval time of our algorithm does not depend on the sizes of the databases. The retrieval time complexities and execution times (specific to computing environment Intel Core2Duo, 2.0-GHz processor, 2.0-GB Memory) of different indexing mechanisms are summarized in Table IV. C. Threats to Validity Our experiments are involved with five iris image databases and different parameters. Eventually, the experimental results we have reported in this work are subjected to the validity of the above resources and assumptions on values of parameters. In the following, we discuss the validity of our experiment and experimental results.

TABLE IV RETRIEVAL TIME COMPLEXITY AND EXECUTION TIME INDEXING MECHANISMS

1201

OF

DIFFERENT

Fig. 10. (a) HR and (b) PR with different scales and orientations.

1) Internal Validity: We assess internal validity [52] on the basis of experimental evaluation of the variables. In our experiment, three parameters, namely number of scales, number of orientations, and , are considered. First, we validate the assumed values of number of scales and number of orientations of Gabor filter. Next, we validate the value of . In our experiment, we are to decide the number of scales and orientations for feature extraction that should give better results. It may be noted that if the number of scales and number of orientations increase, then the number of features in the index key increases. On the other hand, the memory requirement to store the index keys also increases when the number of features in

1202

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

cumulative match score. These metrics are commonly used to measure the performance of biometric indexing systems [14], [19], [22]. This way our approach confirms the construct validity. However, metrics such as identification probability [16] and bin miss rate [16] could be of other interest. Since, these are the metrics usually insignificant to establish the claim of efficiency, we ignore these in our work like previous efforts in this area of research. V. CONCLUSION

Fig. 11. (a) HR and (b) PR for different values of .

index key increases. Further, HR increases as the number of features increases. Hence, there is a trade-off between the number of scales and orientations (and hence memory overhead) with the HR. We have experimented with different scales and orientations, and the result on the CASIAV3I database is shown in Fig. 10. In our experiment, we consider ten different scales and ten different orientations. Fig. 10(a) and (b) shows the HR and PR for different combinations of scales and orientations, respectively. We see that a higher number of scales and orientations gives good HR but this also gives high PR which is not desirable. On the other hand, small values of number of scales and orientations give good PR but they give poor HR which is also not desirable. From Fig. 10, we see that the HR does not change significantly when we increase the number of scales and orientations beyond three and four, respectively. Again, from (14), we see that the memory requirement to store the indexed features is directly proportional to the number of features used in indexing. In other words, with large values of number of scales and orientations memory requirement increases greatly without achieving much in HR. With this observation, we have chosen the value of scale as three and orientation as four. In fact, this combination gives a better HR and PR as shown in Fig. 10. We calculate the HR and PR for different values of . Fig. 11 shows the result for different values of . From Fig. 11(a), we observe that the HR does not increase significantly beyond the value of . On the other hand, from Fig. 11(b), we observe that PR is changing rapidly after . Considering this trend, it is appropriate to set the value of as 5. 2) External Validity: We validate the factors which may limit the generalization of experimental results. Our experimental results are based on the Bath, CASIAV3I, CASIAV4T, MMU2, and WVU iris image databases. All these databases are created in a controlled experimental setup, so we should not claim the results applicable to real life applications. Moreover, to establish the results it needs to be validated with other databases, which we could not access during our experiments. 3) Construct Validity: We would also like to assess how well the theories are implemented into actual programs. The performance of our system is measured by three metrics: HR, PR, and

Recently, a number of approaches have been reported to index iris biometric database. We propose a new indexing technique using Gabor wavelet features. Our proposed technique requires a lower number of features compared to the existing approaches; as a consequence, we need less memory to store index data. Further, with a lower number of features, the index space is organized in such a way that we retrieve a set of identity similar to the query in a constant time. We achieve the memory and computation time advantages without compromising the accuracy. In our approach, we achieve approximately 99% cumulative match score. REFERENCES [1] Iris Recognition [Online]. Available: http://en.wikipedia.org/wiki/ Iris_recognition [2] H. Lee, S.-H. Lee, T. Kim, and H. Bahn, “Secure user identification for consumer electronics devices,” IEEE Trans. Consum. Electron., vol. 54, no. 4, pp. 1798–1802, Nov. 2008. [3] A. Noore, “Highly robust biometric smart card design,” IEEE Trans. Consum. Electron., vol. 46, no. 4, pp. 1059–1063, Nov. 2000. [4] D. D. Hwang and I. Verbauwhede, “Design of portable biometric authenticators—Energy, performance, and security tradeoffs,” IEEE Trans. Consum. Electron., vol. 50, no. 4, pp. 1222–1231, Nov. 2004. [5] D.-S. Kim, S.-Y. Lee, B.-S. Kim, S.-C. Lee, and D.-H. Chung, “On the design of an embedded biometric smart card reader,” IEEE Trans. Consum. Electron., vol. 54, no. 2, pp. 573–577, May 2008. [6] P. Corcoran and A. Cucos, “Techniques for securing multimedia content in consumer electronic appliances using biometric signatures,” IEEE Trans. Consum. Electron., vol. 51, no. 2, pp. 545–551, May 2005. [7] P. Corcoran, C. Iancu, F. Callaly, and A. Cucos, “Biometric access control for digital media streams in home networks,” IEEE Trans. Consum. Electron., vol. 53, no. 3, pp. 917–925, Aug. 2007. [8] Using the Iris Recognition Immigration System (IRIS) [Online]. Available: http://www.ukba.homeoffice.gov.uk/travellingtotheuk/Enteringtheuk/usingiris/ [9] Canpass—Air [Online]. Available: http://www.cbsa-asfc.gc.ca/prog/ canpass/canpassair-eng.html [10] The Eyes Have it [Online]. Available: http://www.accessexcellence. org/WN/SU/irisscan.php [11] UIDAI Strategy Overview [Online]. Available: http://uidai.gov.in/ UID_PDF/Front_Page_Articles/Documents/Strategy_Overveiw001.pdf [12] Multipurpose National Identity Card [Online]. Available: http://en. wikipedia.org/wiki/Multipurpose_National_Identity_Card [13] A. Mhatre, S. Palla, S. Chikkerur, and V. Govindaraju, A. K. Jain and N. K. Ratha, Eds., “Efficient search and retrieval in biometric databases,” in Proc. SPIE Defense and Security Symp., Orlando, FL, Mar. 2005, vol. 5779, pp. 265–273. [14] R. Mukherjee and A. Ross, “Indexing iris images,” in Proc. 19th Int. Conf. Pattern Recognition, Florida, Dec. 2008, pp. 1–4. [15] A. Mhatre, S. Chikkerur, and V. Govindaraju, T. Kanade, A. K. Jain, and N. K. Ratha, Eds., “Indexing biometric databases using pyramid technique,” in Proc. 5th Int. Conf. Audio- and Video-Based Biometric Person Authentication (AVBPA 2005), New York, Jul. 2005, vol. 3546, pp. 841–849. [16] H. Mehrotraa, B. Majhia, and P. Gupta, “Robust iris indexing scheme using geometric hashing of sift keypoints,” J. Netw. Comput. Applicat., vol. 33, no. 3, pp. 300–313, 2010.

DEY AND SAMANTA: IRIS DATA INDEXING METHOD USING GABOR ENERGY FEATURES

[17] H. Mehrotra, G. S. Badrinath, B. Majhi, and P. Gupta, S. Ranka, S. Aluru, R. Buyya, Y. C. Chung, S. Dua, A. Grama, S. K. S. Gupta, R. Kumar, and V. V. Phoha, Eds., “Indexing iris biometric database using energy histogram of dct subbands,” in Proc. Int. Conf. Contemporary Computing (IC3-2009), Noida, India, Aug. 2009, vol. 40, pp. 194–204. [18] S. Palla, S. S. Chikkerur, V. Govindaraju, and P. Rudravaram, “Classification and indexing in large biometric databases,” in Proc. Biometrics Consortium Conf., Virginia, Sep. 2004. [19] B. Bhanu and X. Tan, “Fingerprint indexing based on novel features of minutiae triplet,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 5, pp. 616–622, May 2003. [20] G. Bebis, T. Deaconu, and M. Georgiopoulos, “Fingerprint identification using delaunay triangulation,” in Proc. IEEE Int. Conf. Intelligence, Information, and Systems (ICIIS), Bethesda, MD, Oct./Nov. 1999, pp. 452–459. [21] N. B. Puhan and N. Sudha, “Efficient feature matching in a very large iris database for person identification,” in Proc. 34th Annu. Conf. IEEE Industrial Electronics Society, Orlando, FL, Nov. 2008, pp. 1881–1884. [22] N. B. Puhan and N. Sudha, “A novel iris database indexing method using the iris color,” in Proc. 3rd IEEE Conf. on Industrial Electronics and Applications (ICIEA 2008), Singapore, Jun. 2008, pp. 1886–1891. [23] J. G. Daugman, “High confidence visual recognition of persons by a test of statistical independence,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 15, no. 11, pp. 1148–1161, Nov. 1993. [24] L. Ma, T. Tan, Y. Wang, and D. Zhang, “Personal identification based on iris texture analysis,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 12, pp. 1519–1533, Dec. 2003. [25] L. Ma, T. Tan, Y. Wang, and D. Zhang, “Efficient iris recognition by characterizing key local variations,” IEEE Trans. Image Process., vol. 13, no. 6, pp. 739–750, Jun. 2004. [26] J. Kim, S. Cho, J. Choi, and I. R. J. Marks, “Iris recognition using wavelet features,” J. VLSI Signal Process. Syst., vol. 38, no. 2, pp. 147–156, 2004. [27] Y. Zhu, T. Tan, and Y. Wang, “Biometric personal identification based on iris patterns,” in Proc. 15th Int. Conf. Pattern Recognition, Barcelona, Spain, Sep. 2000, vol. 2, pp. 801–804. [28] C. Sanchez-Avila, R. Sanchez-Reillo, and D. de Martin-Roche, “Irisbased biometric recognition using dyadic wavelet transform,” IEEE Aerosp. Electron. Syst. Mag., vol. 17, no. 10, pp. 3–6, Oct. 2002. [29] S. Dey and D. Samanta, “Fast and accurate personal identification based on iris biometric,” Int. J. Biometrics, vol. 2, no. 3, pp. 250–281, 2010. [30] R. Mukherjee, “Indexing Techniques for Fingerprint and Iris Databases,” Master’s, College of Engineering and Mineral Resources at West Virginia University, Morgantown, WV, 2007. [31] A. D. Alexandrov, W. Y. Ma, A. E. Abbadi, and B. S. Manjunath, “Adaptive filtering and indexing for image databases,” in Proc. SPIE Int. Conf. Storage and Retrieval for lmage and Video Databases—III, California, Feb. 1995, vol. SPIE 2420, pp. 12–23. [32] T. Barbu, “Content-based image retrieval using gabor filtering,” in Proc. 20th Int. Workshop on Database and Expert Systems Applications, Linz, Austria, Aug./Sep. 2009, pp. 236–240. [33] R. S. Choras, T. Andrysiak, and M. Choras, “Integrated color, texture and shape information for content-based image retrieval,” Pattern Anal. Applicat., vol. 10, no. 4, pp. 333–343, 2007. [34] J. Han and K.-K. Ma, “Rotation-invariant and scale-invariant gabor features for texture image retrieval,” Image Vis. Comput., vol. 25, no. 9, pp. 1474–1481, 2007. [35] B. S. Manjunath and W. Y. Ma, “Texture features for browsing and retrieval of image data,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, no. 8, pp. 837–842, Aug. 1996. [36] D. Shin, daewon Kim, H. Kim, and S. Park, “An image retrieval technique using rotationally invariant gabor features and a localization method,” in Proc. 2003 Int. Conf. on Multimedia and Expo (ICME ’03), Maryland, Jul. 2003, vol. 2, pp. 701–704. [37] P. Wu, B. S. Manjunanth, S. D. Newsam, and H. D. Shin, “A texture descriptor for image retrieval and browsing,” in Proc. IEEE Workshop on Content-Based Access of Image and Video Libraries (CBAIVL ’99), Fort Collins, CO, Jun. 1999, pp. 3–7.

1203

[38] C. Wolf, W. Kropatsch, H. Bischof, and J.-M. Jolion, “Content based image retrieval using interest points and texture features,” in Proc. 15th Int. Conf. on Pattern Recognition, Barcelona, Spain, Sep. 2000, vol. 4, pp. 234–237. [39] D. Gabor, “Theory of communication,” J. Inst. Elect. Eng.—Part III: Radio and Commun. Eng., vol. 93, no. 26, pp. 429–457, 1946. [40] J. G. Daugman, “Uncertainty relation for resolution in space, spatial frequency, and orientation optimized by two-dimensional visual cortical filters,” J. Opt. Soc. Amer., vol. 2, no. 7, pp. 1160–1169, 1985. [41] J. G. Daugman, “Complete discrete 2D Gabor transforms by neural networks for image analysis and compression,” IEEE Trans. Acoust., Speech, Signal Process., vol. 36, no. 7, pp. 1169–1179, Jul. 1988. [42] S. Dey and D. Samanta, “An efficient approach to iris detection for iris biometric processing,” Int. J. Comput. Applicat. Technol., vol. 35, no. 1, pp. 2–9, 2009. [43] S. Dey and D. Samanta, “An efficient and accurate pupil detection method for iris biometric processing,” Int. J. Comput. Applicat., vol. 32, no. 2, pp. 141–148, 2010. [44] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 2nd ed. Englewood Cliffs, NJ: Prentice-Hall, 2002. [45] Bath Iris Image Database [Online]. Available: http://www.bath.ac.uk/ elec-eng/research/sipg/irisweb/index.html [46] Casia-IrisV3-Interval Iris Image Database [Online]. Available: http:// www.cbsr.ia.ac.cn/IrisDatabase.htm [47] CASIA-IrisV4-Thousand Iris Image Database [Online]. Available: http://www.cbsr.ia.ac.cn/IrisDatabase.htm [48] MMU2 Iris Image Database [Online]. Available: http://pesona.mmu. edu.my/~ccteo/ [49] WVU Iris Database, Multimodal Biometric Dataset Collection, BIOMDATA, Release 1 2010 [Online]. Available: http://citer.wvu.edu/multimodal_biometric_dataset_collection_biomdata_release1 [50] J. Daugman, “Biometric Personal Identification System Based on Iris Analysis,” U.S. Patent (5291560), 1994. [51] D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar, Handbook of Fingerprint Recognition, 2nd ed. New York: Springer, 2009. [52] B. A. Kitchenham, S. L. Pfleeger, D. C. Hoaglin, and J. Rosenberg, “Preliminary guidelines for empirical research in software engineering,” IEEE Trans. Softw. Eng., vol. 28, no. 8, pp. 721–734, Aug. 2002.

Somnath Dey (S’08) received the B.Tech. degree in information technology from the University of Kalyani, in 2004, and the M.S. (by research) degree in information technology from the School of Information Technology, Indian Institute of Technology, Kharagpur, in 2008. Presently, he is working toward the Ph.D. degree at the School of Information Technology, Indian Institute of Technology, Kharagpur. His research interests include biometrics, image processing, pattern recognition, and human computer interaction.

Debasis Samanta (A’95–S’02–M’02) received the B.Tech. in computer science and engineering from Calcutta University, the M. Tech. degree in computer science and engineering from Jadavpur University, and Ph.D. degree in computer science and engineering from Indian Institute of Technology, Kharagpur. He is currently an Associate Professor in the School of Information Technology, Indian Institute of Technology, Kharagpur. His research interests include human computer interaction, biometric processing, and software testing.