This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Application of Pyramidal Directional Filters for
Biometric Identification using Conjunctival
Vasculature Patterns Sriram Pavan Tankasala, Plamen Doynov, and Reza Derakhshani
Department of Computer Science Electrical Engineering, University of Missouri – Kansas City
454 Robert H. Flarsheim Hall, 5110 Rockhill Road, Kansas City, MO 64110-2499, US
Abstract— Directional pyramidal filter banks as feature
extractors for ocular vascular biometrics are proposed. Apart
from the red, green, and blue (RGB) format, we analyze the
significance of using HSV, YCbCr, and layer combinations
(R+Cr)/2, (G+Cr)/2, (B+Cr)/2. For classification, Linear
Discriminant Analysis (LDA) is used. We outline the advantages
of a Contourlet transform implementation for eye vein
biometrics, based on vascular patterns seen on the white of the
eye. The performance of the proposed algorithm is evaluated
using Receiver Operating Characteristic (ROC) curves. Area
under the curve (AUC), equal error rate (EER), and decidability
values are used as performance metrics. The dataset consists of
more than 1600 still images and video frames acquired in two
separate sessions from 40 subjects. All images were captured
from a distance of 5 feet using a DSLR camera with an attached
white LED light source. We evaluate and discuss the results of
cross matching features extracted from still images and video
recordings of conjunctival vasculature patterns. The best AUC
value of 0.9999 with an EER of 0.064% resulted from using Cb
layer in YCbCr color space. The best (lowest value) EER of
0.032% was obtained with an AUC value of 0.9998 using the
Within the past decade, biometric-based personal authentication technologies have found many applications [1, 2]. The historical dominance of face and fingerprint modalities is being challenged by other biometric modalities such as iris, vein patterns, palm prints, hand geometry, voice, and DNA short tandem repeats [3]. The success of ocular biometrics is based on its inherent advantages [4] and recent progress in related supporting technologies and processing algorithms [5-11]. However, many challenges remain, especially with respect to variable image acquisition conditions and the required degree of user cooperation. Retina imaging techniques require specialized devices, very close proximity, and user cooperation [12-14]. Iris recognition is relatively well established and accepted. However, the iris modality requires near infrared imaging for the majority of dark, pigment-rich eyes. Recently, personal recognition using ocular imaging in the visible spectrum has received increased attention. This is especially true for vasculature seen on the white of the eye. Images of the eye in the visible spectrum reveal the vascularity of the outer coatings of the eyes (mostly due to conjunctival and episcleral layers, Figure 1). The conjunctiva and its underlying episclera
are anterior segment structures of the human eye, exposed to the naked eye and easy to capture with regular RGB cameras. The covering mucous membrane is clear and facilitates the imaging of the vasculature on the outer surface of the bulbar layer. In interest of brevity, we will henceforth refer to the plurality of conjunctival and episcleral vasculature seen on the white of the eye as conjunctival vasculatures (CV). Conjunctival tissue lines the inside of eyelids and spreads over the anterior sclera (white part of the eye) up to the scleral-corneal limbus. The diversity of CV during pattern formation provides an immense amount of unique textural information, which can be used as biometric tokens.
Figure 1: Wavelet vs. Contourlet approach across vascular patterns
The conjunctival vasculature can be used as a separate modality or to complement the iris modality to compensate for iris images with off-angle gaze (especially images captured at an extreme gaze in the left or right direction). Previous work on textural classification of conjunctival vasculature has demonstrated high accuracies that support the practical use of conjunctival vasculature as a biometric modality [7-11, 15-17]. In general, biometrics is a pattern recognition problem, and thus is heavily dependent on the following two stages: feature extraction and classification (matching). The performance of any biometric system depends on the reliable and robust feature extraction. Previous work on conjunctival vasculature recognition shows the importance of various feature extraction methods for obtaining higher accuracies [7, 11, 15-17]. In this paper, we propose and evaluate the performance of methods that use HSV and YCbCr color presentation before feature
extraction and thus avoiding the non-uniformity of the RGB color space. The features are extracted locally and globally from color and texture information. We examine the feature extraction using different color models and demonstrate the differences. We used Pyramidal directional filtering approach (Contourlets) for feature extraction. Contourlets overcome the limitations of traditional wavelets [18, 19]. As an extension of wavelets with an added property of multi-directionality, the Contourlet transform has the ability to extract edge information as well as smooth contour information. Contourlets are the discrete version of the Curvelets with added benefits of multi-resolution and multi-directional functionality. Similar to 2D-Curvelets, the Contourlets have advantages to process edges as curves and derive reliable information from image patterns. Encouraged by the successful use of Contourlets in various image processing applications [18], we propose algorithms based on Contourlets for feature extraction in ocular biometrics. The rest of this paper is organized as follows. Section II
describes the data collection protocol, the image
preprocessing, and the segmentation of scleral regions of
interest (ROI). Section III provides background on Contourlets
with some details and formulas for the Pyramidal direction
approach in Contourlets, as well as the classification methods
used. Section IV presents the implementation details. The
results are presented in Section V and are followed by
discussions in Section VI. The final section VII provides a
summary of the main results along with conclusions and
directions for future work.
II. DATA COLLECTION PROTOCOL, SEGMENTATION AND
PREPROCESSING
A. Data collection protocol
Data was collected from 40 volunteers (IRB protocol 11-57e). For image acquisition, we used a Canon T2i DSLR camera with an attached Digi-Slave Flex 6400 white LED illumination source, which is a macro ring light with two extended lateral lighting pads. Eye images were captured in still and video mode. In still mode, the camera was operating in its native burst speed of 4 frames per second (fps). In 1080 progressive scan video mode, the capture rate was 30 fps. Data was collected in two consecutive sessions separated by a 30-minute rest interval. Data from session I were used for training and those of session II were used for testing. Subjects were asked to avert their gaze to left and right for maximum exposure of the sclera and CV (Figure 8). The average distance from the camera to the subject was kept at 5 feet. The camera was equipped with a Canon EF 70mm zoom lens and was operated at f-stop 5.6, an exposure time of 1/125 sec, and ISO 800. For each subject, five samples (from the multiple still images) and five frames (from 1080p video) were randomly chosen from session-I and session-II for further processing.
B. Segmentation of sclera
Segmentation of scleral regions of RGB ocular images was
performed using k-means clustering with Euclidean distance
metric and k=3 [11]. Each pixel is denoted as a 3-dimensional
vector representing its RGB intensities. The pixels pertaining
to the scleral region were determined as the cluster with the
largest Euclidean distance from the origin of the coordinate
system to its centroid. The pixels belonging to the iris region
(the iris mask) were determined as the cluster with the
smallest Euclidean distance from the origin of the coordinate
system to its centroid. See Figure 2. The largest connected
region was selected for scleral and iridial masks. Due to the
presence of artifacts such as specular reflections and glare,
some pixels within the sclera region were not assigned to the
proper cluster, thereby appearing as holes in the sclera mask.
To smooth the contour of the scleral mask and to fill its
aforesaid voids, a convex hull operation is applied to the
clustering-derived scleral mask. The convex hull operation
may incorrectly include some pixels pertaining to the iris
cluster and adjacent to the limbic boundary in the scleral
mask. To address this problem, we remove the convex hull of
the iridial mask from the convex hull of the scleral mask. In
the end, for each eye image, a maximum area rectangle
inscribed in the scleral mask is extracted from the scleral
region, as shown in Figure 2, and is designated as the region
of interest (ROI) for the forthcoming operations.
Figure 2: Segmentation of sclera using k-means clustering algorithm.
a) Original image b) Resultant sclera mask c) Sclera mask imposed on the
original image d) Inscribed max area rectangle on segmented sclera
C. Preprocessing
A Contrast Limited Adaptive Histogram Enhancement
algorithm (CLAHE) was used to enhance vascular patterns of
the ROI. The CLAHE algorithm is performed on non-
overlapping partitions (tiles) of an image (8 x 8 tiles per ROI
in this study). The contrast of the each tile is enhanced in such
a way that its histogram matches a specified histogram shape
(flat histogram in this study). Bilinear interpolation is applied
to eliminate edge effect at the boundaries of the tiles at the
time of reconstruction [20]. RGB, YCbCr, and HSV color spaces and exponential
power density function: HSV presents color and intensity
values in a more intuitive and perceptually relevant fashion in
comparison to the Cartesian (cube) representation. The hue
(H) and saturation (S) components of the HSV color space are
640
intimately related to the way the human eye perceives colors.
The value (V) relates to the intensity of the colors. Because
HSV is a simple transformation of the device-dependent RGB
model, its defined physical colors depend on the RGB
primaries of the device (or on the particular RGB profile it
represents), including gamma corrections. Therefore, each
unique and device dependent RGB profile has a unique HSV
space.
YCbCr is another color representation used for color
processing and perceptual uniformity. Its luminance is very
similar to the grayscale version of the original image. A
Chrominance-Blue (Cb) channel is strong in parts of the
image containing blue color. Similarly, Chrominance-Red (Cr)
is strong in places of occurrence of reddish colors, and both
Cb and Cr are weak in green color regions. For the color space
transformations and for computations presented in this paper,
we used the MATLAB 2012b (MathWorks, MA) along its
Image Processing and Parallel Computing toolboxes.
An exponential power density function applies an
exponential distribution to an intensity image, and is described
as
( ⁄ )
The above formula, in which is the mean parameter, was
applied to each pixel of the ROI. As a result, higher intensities
are attenuated, accentuating the lower intensities of the ROI
vascular patterns (Figure 4).
III. THEORY ON CONTOURLETS AND CLASSIFICATION
A. Introduction to Contourlets
Contourlets apply a two-step filter bank to extract
information from contour-rich patterns of an image. The
Contourlet transform was first introduced by Do and Vitterli in
2001 [19]. Contourlets were developed for extracting reliable
information from the contour pattern segments of an image
and overcoming the limitations of traditional wavelets in this
regard [19]. The Laplacian pyramid, in conjunction with
directional filter bank, are used as a two-step filter bank.
B. Laplacian pyramids
The Laplacian pyramid decomposes an image into low-pass and band-pass images (Figure 3). The process generates a low-pass version of the original image (LPout) and band-pass filtered image (BPout). BPout is difference between original image and synthesis filtered image (SF) [21,22].
A directional filter bank was efficiently implemented using
an l-level binary tree decomposition which results in 2l sub-
bands that use a wedge shaped filter. Decomposition tree
expansion includes two building blocks. The first block
involves a two channel Quincunx filter bank for dividing the
2D-spectrum into horizontal and vertical directions. Quincunx
filter banks are two-dimensional, non-separable filter banks
and are widely used in many signal processing applications
[23]. The second block is a shear operator [19]. From the
above Laplacian pyramid and directional filter banks, the
Contourlet transform is given as
∑
[ ] (2)
In the above equation,
[ ] is the directional
filter bank basis, and is the Laplacian pyramid, l is
the level of decomposition,
is the sampling matrix, and
is the synthesis filter.
D. Classification
Linear Discriminant Analysis (LDA) is a supervised linear classification and dimensionality reduction method for casting multi-dimensional features into a single dimension in a way that the projected data points, of the original classes, are maximally separable. In this study, Fisher’s LDA was used [24].
E. Performance metrics
The receiver operating characteristic (ROC) analysis was
used to test the performance of the classifier [25]. We used the
Area under the ROC curve (AUC), Equal error rate (EER),
and decidability distance as performance metrics. The
decidability index (d') was introduced by Daugman for
evaluating the quality of typical dichotomies (two-class
separability ).
F. Match-score level fusion
Various fusion techniques can be applied to combine
information for better classification. In general, match score
level fusion is a better choice to improve the classifiability and
to lower the errors given multiple biometric cues. Typical
techniques used in match score level fusion are the simple (but
yet effective) sum rule, min rule, and max rule [26].
IV. EXPERIMENTAL PROCEDURE
A. Segmentation of sclera and preprocessing
To verify proper segmentation, results were visually inspected and corrected when necessary. After k-means segmentation of the sclera, the average initial size of the ROIs (max-area rectangles inscribed in the segmented sclera) for still images was close to 256x256 pixels. The average initial size of the ROIs obtained from video-frame images was close to 128x128 pixels. Thus, the images were resized to the aforesaid powers of two. In order to compare the influence of image resolution on the results, downsampled and non-downsampled versions were processed.
641
B. Preprocessing
CLAHE was applied to the ROIs with the following
parameters: 8×8 tiles, contrast enhancement limit of 0.01, and
uniform histogram. Consecutively, the images were further
processed using an exponential power density function for
vessel enhancement with a mean value of 0.25, as shown in
Figure 4.
Figure 4: Original green color layer of max-area rectangle ROI (left), the same
image after CLAHE application (center), and after enhancement with
exponential power density function (right)
C. Feature extraction
Feature extraction was performed using the Contourlet
transform [19]. Preprocessed ROIs were used as an input to
the Contourlet transform with the level of decomposition l = 4.
For each sub-band of the decomposed image, a feature vector
was formed using the calculated mean, median, variance, and
entropy. For the current study, multiple separate layers from
different color models or color layers in some combination
were used. In this paper, we present the results from using
separate layers from the RGB, HSV, YCbCr color spaces and
a combination of layers (R+Cr)/2, (G+Cr)/2, (B+Cr)/2.
D. Classification
One-versus-rest Fisher’s LDAs were trained with
Contourlet features from still mode and video mode images in
session I (one to one biometric verification analysis). We
tested the trained LDA with images from session II. Cross
matching between the still mode and video mode images and
vice versa was also performed. This was repeated for the four
possible image capturing scenarios: two gaze directions (left
and right), for each eye (left and right). We use the
abbreviated notations left eye looking left (LLL), left eye
looking right (LLR) for the left eye, and similarly RLL and
RLR for the right eye. The cross matching was performed for
all possible capturing scenarios using the statistical parameters
(mean, median, variance, and entropy) of the Contourlet
transform coefficients of each sub-band. We performed match
score fusion using three techniques (sum rule, min rule, and
max rule) over four ROI’s (LLL, LLR, RLL, and RLR) across
all image sizes of each ROI. ROC analysis was performed to
evaluate the performance of the LDA classifier [21]. AUC and
EER of the ROCs and d’ values were reported for individual
layers from RGB, HSV, YCbCr color spaces, and for the
(R+Cr)/2, (G+Cr)/2, (B+Cr)/2 combinations.
V. RESULTS
A. Results using Fisher’s LDA with still images
We calculated the ROC AUC’s, EER’s and d’ values for each color space using Contourlet features for still images. The best AUC value of 0.9999, with an EER of 0.064% and a d’ value of 3.728, resulted using min rule for the Cb layer in the YCbCr color space. The best (lowest value) EER of 0.032% was obtained with an AUC value of 0.9998 and a d’ value of 4.197 using min rule and green layers of the original RGB images. The best (maximum) d’ value of 4.572 and an associated AUC of 0.9930 and EER of 2.4% were obtained using the green layer in the RGB images and simple sum rule
for match score level fusion.
B. Results using Fisher’s LDA with video frames
AUC’s, EER’s and d’ values for the each color space were calculated using video frames. The highest AUC value of 0.9970 with an EER of 2.5% and d’ value of 3.156 was obtained using min rule with green layer of the RGB image.
C. Results using Fisher LDA with still images for training and
video frames for testing
The highest AUC value of 0.9990 with an EER of 0.16% resulted using min rule with the green layer of an RGB image. The performance of the green + Cr color combination placed second with an EER of 0.67% and AUC of 0.9982. The highest d’ value of 3.636 (high inter-class separability) in the still images/video frames cross matching was obtained with the blue layer in the RGB image using the simple sum rule for match score level matching.
D. Results using Fisher LDA with video frames for training
and still images for testing
In this scenario, video frames were used for training the
LDA (enrollment) and still frames were used for testing. The
best AUC value of 0.9927 with an EER value of 2.6% and a d’
value of 2.673 were obtained using the min rule with V (value
layer) in the HSV color space. The green + Cr combination
performed slightly lower, with an AUC of 0.9900 and an EER
value of 4.8%, but with a higher d’ value of 3.410. Figures 5
and 6 show the EER’s and d’ results for all color transformed
still images, video frames, still images vs. video frames, and
video frames vs. still images.
VI. DISCUSSION
The CLAHE method was used in most of the previous work for vessel enhancement for CV recognition [7, 10, 11]. In this study, in conjunction with CLAHE, we used an exponential power density function as shown in Figure 4. The visibility of the texture is better compared to the enhancement with CLAHE only. The AUC, EER, and d’ values for the images enhanced with an exponential power density function are also better when compared to CLAHE preprocessing only. The summary of results for all the three performance factors is shown in in Table 1. As can be seen, the AUC increases from 0.9277 to 0.9938 and the EER decreases from 3.685% to 2.40 %, and d’ increases from 2.7247 to 4.5697, supporting the advantage of exponential enhancement.
642
TABLE1. AUC, EER, AND D’ VALUES FOR CLAHE AND CLAHE + EXPONENTIAL POWER DENSITY FUNCTION ENHANCEMENT OF THE GREEN LAYER OF RGB IMAGE USING STILL IMAGES WITH SIMPLE SUM RULE MATCH SCORE FUSION.
AUC's EER's [%] d’
CLAHE 0.9277 3.6859 2.7247
CLAHE + Exponential
enhancement
0.9938 2.4038 4.5697
The process of cross matching between still images and videos is performed to investigate the reliability and robustness of the Contourlet features with various color transformed images across different image capturing methods for CV recognition. Though acquired with the same camera, the resolution and compression of the still images and the video frames are different. The results indicate that for a practical implementation, still images need to be used for enrollment and video frames could be used for verification (Figure 7). In the current study, the images were randomly chosen from each capture stack (sequence of frames taken in a short period of time) for further processing and matching. To achieve better performance, the random selection of video frames has to be replaced with certain frame selection criteria (image quality). Previous work on CV mostly includes usage of the green
layer of the RGB images [7-11]. In this study, we additionally
used HSV and YCbCr color transformed images. We also
investigated red and blue layers. The results clearly show that
the green layer performed better in most of the cases. Within
the confines of our dataset, the value V of HSV worked better
than the green layer with an EER of 2.6% for videos vs. still
images. The Cb layer from YCbCr yielded an AUC of 0.9999
with an EER of 0.064 % for still images. The lowest EER of
0.032% was obtained for RGB stills using their green layer. In
summary, RGB’s green layer performed the best overall. The V
layer from HSV, Cr and Cb from YCbCr, blue from RGB, and
green + Cr layer combination may also be used to obtain
texture information. The aforesaid can be visually inferred from
the results presented in Figures 5 and 6.
Figure 5: EER’s (lower is better) for stills, video frames, stills vs. video
frames, and video frames vs. stills; using LDA with min rule match score
level fusion.
Figure 6: d’ (higher is better) for stills, video frames, still vs. video frames,
and video frames vs. stills; using LDA with min rule match score level fusion.
Figure 7: ROC curves for best results using still image, video frames, still vs.
video, and video vs. stills.
a)
b)
Figure 8: Sample eye images from a) still and b) video frames database used in the study
643
VII. CONCLUSION AND FUTURE WORK
Contourlet based CV biometric recognition was
successfully performed with a best EER of 0.032% for still
images using green layer of RGB captures. Image
enhancement for better texture representation using CLAHE +
exponential power density function was successfully
implemented. From the results we conclude that, apart from
green layer of RGB color space, HSV and YCbCr color spaces
can also be used for CV biometric recognition.
For future work, we would like to investigate the
performance of the proposed method over larger datasets. We
would also like to study different fusion techniques using
wavelets, Contourlets and gray level co-occurrence matrices to
build a more robust multi-algorithmic CV biometric
recognition system
ACKNOWLEDGEMENT
Research was sponsored by the Leonard Wood Institute in
cooperation with the U.S. Army Research Laboratory and was
accomplished under Cooperative Agreement Number
W911NF-07-2-0062. The views and conclusions contained in
this document are those of the authors and should not be
interpreted as representing the official policies, either
expressed or implied, of the Leonard Wood Institute, the
Army Research Laboratory or the U.S. Government. The U.S.
Government is authorized to reproduce and distribute reprints
for Government purposes notwithstanding any copyright
notation heron. Authors thank Mr. Sashi Saripalle for his help
in data collection. Authors would like to thanks Dr. Simona
Crihalmeanu and Dr. Arun Ross from Michigan State
University for providing scleral segmentation algorithm and
“Biometrics of Next Generation: An Overview"," 2010. [2] D. Bhattacharyya, R. Ranjan, P. Das, K. Tai-hoon, and S. K.
Bandyopadhyay, "Biometric Authentication Techniques and its Future
Possibilities," in Second International Conference on Computer and Electrical Engineering, 2009. ICCEE '09, 2009, pp. 652-655.
[3] L. Nadel and T. Cushing, "Eval-Ware: Biometrics Resources [Best of the
Web]," IEEE Signal Processing Magazine , vol. 24, pp. 136-139, 2007. [4] A. K. Jain, A. Ross, and S. Prabhakar, "An introduction to biometric
recognition," IEEE Transactions on Circuits and Systems for Video
Technology, , vol. 14, pp. 4-20, 2004.
[5] K. W. Bowyer, K. Hollingsworth, and P. J. Flynn, "Image understanding for
iris biometrics: A survey," Computer Vision and Image Understanding, vol.
110, pp. 281-307, 2008.
[6] J. Daugman, "New Methods in Iris Recognition," IEEE Transactions on
Systems, Man, and Cybernetics, Part B: Cybernetics, , vol. 37, pp. 1167-1175, 2007.
[7] R. Derakhshani and A. Ross, "A Texture-Based Neural Network Classifier
for Biometric Identification using Ocular Surface Vasculature," in International Joint Conference on Neural Networks, 2007, pp. 2982-2987.
[8] V. Gottemukkula, S. K. Saripalle, S. P. Tankasala, R. Derakhshani, R.
Pasula, and A. Ross, "Fusing iris and conjunctival vasculature: Ocular biometrics in the visible spectrum," in IEEE Conference on Technologies for
Homeland Security (HST), 2012, 2012, pp. 150-155.
[9] R. Derakhshani, A. Ross and S. Crihalmeanu, "A New Biometric Modality Based On Conjunctival Vasculature," Proc. of Artificial Neural Networks in
Engineering (ANNIE), (St. Louis, USA), November 2006.
[10] S. P. Tankasala, P. Doynov, R. R. Derakhshani, A. Ross, and S. Crihalmeanu, "Biometric recognition of conjunctival vasculature using GLCM
features," in International Conference on Image Information Processing
(ICIIP), 2011, 2011, pp. 1-6. [11] M. Tistarelli, M. Nixon, S. Crihalmeanu, A. Ross, and R. Derakhshani,
"Enhancement and Registration Schemes for Matching Conjunctival
Vasculature," in Advances in Biometrics. vol. 5558, ed: Springer Berlin / Heidelberg, 2009, pp. 1240-1249.
[12] C. Fancourt, L. Bogoni, K. Hanna, Y. Guo, R. Wildes, N. Takahashi, and
U. Jain, "Iris Recognition at a Distance," in Audio- and Video-Based Biometric Person Authentication. vol. 3546, Springer Berlin Heidelberg, 2005, pp. 1-13.
[13] S. Marios.Unconstrained Biometric Identification: Emerging
Technologies. [14] A.Ross, "Recent Progress in Ocular and Face Biometrics: A CITeR
Perspective," 2010. [15] O. Kangrok and T. Kar-Ann, " Extracting sclera features for
cancelableidentity verification," in International Conference on Biometrics
(ICB), 2012 5th IAPR, 2012, pp. 245-250. [16] N. L. Thomas, Y. Du, and Z. Zhou, "A new approach for sclera vein
recognition," pp. 770805-770805, 2010.
[17] Z. Zhou, Y. Du, N. L. Thomas, and E. J. Delp, "Multimodal eye recognition," pp. 770806-770806, 2010.
[18] A. Azizi and H. Pourreza, "A Novel Method Using Contourlet to Extract
Features for Iris Recognition System," in Emerging Intelligent Computing Technology and Applications. vol. 5754, Springer Berlin Heidelberg, 2009, pp.
544-554.
[19] M. N. Do and M. Vetterli, "The contourlet transform: an efficient directional multiresolution image representation," IEEE Transactions on Image
Processing, , vol. 14, pp. 2091-2106, 2005.
[20] K. Zuiderveld, "Contrast limited adaptive histogram equalization," in Graphics gems IV, ed: Academic Press Professional, Inc., 1994, pp. 474-485.
[21] M. N. Do and M. Vetterli, "Framing pyramids," IEEE Transactions on
Signal Processing, , vol. 51, pp. 2329-2342, 2003. [22] P. J. Burt and E. H. Adelson, "The Laplacian Pyramid as a Compact Image
Code," IEEE Transactions on Communications, vol. 31, pp. 532-540, 1983.
[23] M. Vetterli, "Multi-dimensional sub-band coding: Some theory and algorithms," Signal Processing, vol. 6, pp. 97-112, 1984.
[24] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification (2nd
Edition): Wiley-Interscience, 2000. [25] T. Fawcett, "An introduction to ROC analysis," Pattern Recognition
Letters,Elsevier, vol. 27, pp. 861-874, 2006.
[26] A. Ross and A. Jain, "Information fusion in biometrics," Pattern Recognition Letters, vol. 24, p. 11p, 2003.