You are on page 1of 5

IJCB IJCB Submission ****. CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.

IJCB
#**** #****

100 150
101 151
102 Impact of Resolution and Blur on Iris Identification 152
103 153
104 Anonymous IJCB 2011 submission 154
105 155
106 Paper ID 298 156
107 157
108 158
109 measurement), performance (accuracy, speed, and 159
110 Abstract robustness of the technology used), acceptability (degree of 160
111 approval of a technology), and circumvention (ease of use 161
112 Iris recognition is increasingly employed as a biometric of a substitute for the characteristic) [2]. 162
113 modality to improve the security across the government, 163
114 public, and private sectors; however, the limitations of iris Irises are perceived to be one of the best characteristics 164
115 recognition are yet to be fully understood. Many algorithms to use as a biometric trait because the error rate of iris 165
116 have been proposed which provide several options for recognition is one of the lowest among known biometric 166
117 creating a system to identify an iris. All of these algorithms traits [3]. Everyone has two irises from birth until death. No 167
118 have at least one common weakness, obtaining ideal iris two irises are considered to be the same [3]. Irises can 168
119 images. All iris identification systems are limited in their contain many distinctive features such as arching 169
120 ability to correctly identify an iris by imperfectly acquired ligaments, furrows, ridges, crypts, coronas, freckles, and 170
121 target images. Two challenges that must be considered zigzag collarets [4]. The iris has the great mathematical 171
when acquiring target images are resolution and blur. The advantage that its pattern variability among different
122 172
impact of these two challenges on correct recognition rates persons is enormous [5]. Irises remain stable from six
123 173
have been explored and summarized in this work in an months of age until death [6]. Since the iris is an internal
124 organ that can be seen externally, iris identification systems
174
125 attempt to provide clarity for challenges related to iris 175
biometrics. can be noninvasive [5]. In recent decades, technology has
126 finally been able to answer academia's call for an iris 176
127 identification system. Burch's initial proposal of using the 177
128 1. Introduction iris as a pattern for identification has become a reality [6]. 178
129 As the level of security breaches and transaction fraud The iris pattern is considered to be difficult to reproduce 179
130 increases, the need for highly secure identification and [7], however even commercial iris recognition systems 180
131 personal verification technologies is becoming apparent. have frequently accepted reproduced irises as legitimate 181
132 This need for biometrics can be found in federal, state, and irises [8]. 182
133 local governments, in the military, and in commercial 183
134 applications. It is clear that every entity which provides Regardless of the contemporary research on iris 184
135 access to private information can greatly benefit from a recognition, important questions regarding robustness of 185
136 secure identification and personal verification system. iris identification in various usability scenarios remain. 186
137 Among those there are questions concerning the limitations 187
138 Many systems have been proposed as a solution for such of iris identification for cases of extreme degradation of 188
139 a need including but not limited to fingerprinting, palm image quality. The objective of this paper is to provide a 189
printing, vascular pattern recognition, hand geometry, better understanding of the impact of captured iris
140 190
dynamic signature, voice recognition, facial recognition, dimensions and image blur on corresponding identification
141 191
and iris identification [1]. accuracy.
142 192
143 193
144
1.1. Motivation of Iris Identification 1.2. Related Work 194
145 In order for a human characteristic to be used as a Many solutions have been proposed for answering the 195
146 biometric, it must meet specific criteria. Any potential demand for an iris identification system. Improvements 196
147 biometric characteristic is analyzed against the biometric have been proposed to these systems in an attempt to 197
148 rubric. This rubric consists of qualifiers: universality (each optimize their performance. However, precious little is 198
149 person should have the characteristic), uniqueness (how written concerning some of the limitations associated with 199
well the characteristic separates individuals), permanence all iris identification systems, namely image resolution and
(how well the characteristic resists aging and other time image blur.
dependent variances), collectability (ease of acquisition for

1
IJCB IJCB Submission ****. CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE. IJCB
#**** #****

200 Although the non invasive nature and high variability of 3.2. Image Resolution 250
201 irises make them a promising biometric, high quality 251
photographs with high resolution and good contrast are Our goal was to test iris image resolution requirements.
202 252
needed for todays iris identifying algorithms to perform In order to define the effect of image resolution on iris
203 recognition, several datasets of varying iris image
253
204 well. This is not a problem if the iris is photographed at 254
very close range with a cooperative subject, but if the resolutions were created. The initial dataset was simply the
205 raw first session of Ubiris3 of UBIRIS.v1. This dataset was 255
206 subject is farther away and possibly even walking, 256
photographing the iris for identification becomes more then copied twelve times. The resolution of each copy's iris
207 images were then decreased. Microsoft Office Picture 257
208 challenging [9]. 258
Manager was used to alter image resolution. Microsoft
209 Office Picture Manager allows the resolution of multiple 259
According to Daugman, a minimum radius of 70 pixels
210 images in multiple folders to be changed simultaneously. 260
(140 pixel diameter) is required in order to capture the rich
211 details of the iris pattern [10]. However, ISO/IEC has set
261
212 the standard of a required minimum 200 pixel diameter 262
213 across the iris for high quality iris images [10].
A B 263
214 Hollingsworth's research in how dilation of the pupil 264
215 affects biometric performance suggests that the diameter of 265
216 the iris is an inadequate way to measure the total amount of 266
217 iris available. She asserts that the annular width of the iris is 267
218 a more correct measure of the iris size [11]. 268
219 269
220 Algorithms have been proposed to de-blur images [12]. 270
221 These algorithms are employed when an image is found to 271
222 be sufficiently blurry, cannot be re-acquired, and whose 272
223 correct iris identification is less than sufficient. Some 273
224 proposals are designed to select the best quality iris image C 274
225 from a set of images [13]. Studying the effect of severe 275
226 image compression on iris recognition performance has 276
227 been pursued [14]. The compressed images may be 277
228 perceived to be blurred although there is a significant 278
229 difference between a blurred image and a compressed 279
image. To the best of our knowledge, no one has ever
230 280
attempted to investigate the relationship between image
231 281
blur and correct iris identification. Figure 1 : A) 1% resolution B) 50% resolution
232 282
233 C) 100% resolution 283
234 2. Objectives The above iris resolutions were created starting with a 284
235 The objective of this research is to investigate the very small resolution and reaching maximum available 285
236 relationship between: 1) Image Resolution and size. Table 1 presents tested resolution levels. 286
237 Identification Performance and 2) Image Blur and 287
238 Identification Performance. Iris Sample 288
Image Image
239 Annular Iris Pixels 289
Width Height
3. Methodology Width Diameter
240 290
3 8 20 15 300
241 Four components are required to test the effect of the iris 291
7 20 45 34 1530
242 size and image blur on iris recognition rates: 1) An iris 292
10 28 63 47 2961
243 database had to be acquired 2) The ability to alter the image 14 40 89 67 5963
293
244 resolution 3) The ability to alter image blur 4) Iris 17 49 110 82 9020
294
245 identification software. 19 56 126 95 11970 295
246 22 63 141 106 14946 296
247 3.1. Iris Database 24 69 155 116 17980 297
248 26 75 167 125 20875 298
The UBIRIS database was selected among other
249 databases (CASIA, MMU, etc.) because it provides high 28 80 179 134 23986 299
quality iris images [15] with minimized noise factors and is 29 85 190 142 26980
widely accepted among iris recognition researchers. 31 90 200 150 30000

2
IJCB IJCB Submission ****. CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE. IJCB
#**** #****

300 Table 1: Iris Image Resolution Information 25 29.41% 350


301 33 38.82% 351
302 3.3. Image Blur 39 45.88% 352
303 43 50.59% 353
Our goal was to blur the image in the same way as done 47 55.29%
304 by an out of focus camera. In photography, bokeh is the
354
305 51 60.00% 355
blur or aesthetic quality of the blur [16, 17]. Bokeh can be 59 69.41%
306 simulated by convolving the image with a kernel that 356
307 61 71.76% 357
corresponds to the image of an out-of-focus point source 65 76.47%
308 taken with a real camera. Gaussian blur is less 358
67 78.82%
309 computationally expensive and produces a softer effect 359
71 83.53%
310 than convolution. 360
77 90.59%
311 81 95.29%
361
312 The Gaussian blur filter uses a Gaussian function for 83 97.65% 362
313 computing the alteration applied to each pixel. The 85 100.00% 363
314 equations of a Gaussian function in one and two Table 2: Iris Image Blur Information 364
315 dimensions: 365
316 366
317 Equation 1: 1-Dimensional 367
318 1
 A 368
 = 
319 2
369
320 370
321 Equation 2: 2-Dimensional 371
322 1     372

,  =  
323 2
373
324 374
325 In order to apply a Gaussian function to an image, a C++ 375
326 program was developed that can apply a Gaussian blur to 376
327 multiple images in multiple directories. Librow's Gaussian 377
328 filter implementation was used [18]. This filter is based on 378
329 a window. As the window size increases, so does the blur. B 379
330 380
To produce the largest possible blur for our database, we
331 381
searched for the lowest window size corresponding to a
332 382
CRR of 0%. We found that the lowest window size which
333 produced a CRR of 0% was 85 (=85). The window size of
383
334 0 represents 0% blur which corresponds to the original 384
335 dataset. Twenty-six unique window sizes were selected 385
336 between those extreme points to represent the whole range 386
337 of possible blur levels. A dataset was created for each of the 387
338 26 selected window sizes. Table 2 presents tested blur 388
339 levels. Figure 2: A) Window size 9 389
340 B) Window size 85 390
341 Window Size () % Blur 391
342 0 0.00% 3.4. Iris Recognition Software 392
343 1 1.18% 393
The software suite that was selected was GIRIST, a free
344 3 3.53% 394
5 5.88% iris recognition system [19]. GIRIST performs comparably
345 395
7 8.24% to the best commercial systems deployed today [23].
346 GIRIST computes Correct Recognition Rate (CRR) which
396
347 9 10.59% 397
11 12.94% corresponds to true accept rate [20] of a biometric system.
348 GIRIST produces 96.8% CRR on average when tested on 398
349 13 15.29% 399
15 17.65% the following databases: CASIA-IrisV3-Interval [21],
17 20.00% UBIRIS_1 [15], MMU iris [22].
21 24.71%

3
IJCB IJCB Submission ****. CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE. IJCB
#**** #****

400 4. Results 450


401 451
402 4.1. Resolution Impact Figure 4 shows that as blur increases, CRR decreases. 452
403 CRR only drops ~1.5% as blur increases from 0% to 5%. 453
404 Figure 3 presents CRR results for the iris resolutions The recognition rate decreases linearly as blur increases 454
405 represented by Table 1. from 10% to 30% (~10% blur ~90% CRR, ~20% blur 455
406 ~80% CRR, ~30% blur ~70% CRR). A big drop occurs 456
407 near 30% blur. Unfortunately, this drop is not extremely 457
408 significant because the CRR has already decreased 458
substantially and therefore the behavior past this point is
409 459
trivial.
410 460
411 This graph demonstrates very clearly that it is possible
461
412 for researchers to sacrifice a small amount of blur on the 462
413 entire image in an attempt to reduce the image's high 463
414 frequency components. 464
415 465
416 5. Limitations 466
417 467
418 468
5.1. Image Blur
419 469
420 Figure 3: Correct Recognition Rate as a function of Iris The Gaussian blur is an emulation of depth of field blur 470
421 Diameter found in common photographs, and thus the inferences we 471
422 may make from this research may not pertain to all blur 472
423 It is clear that as resolution decreases, CRR decreases. situations. Therefore further research is necessary to define 473
424 The perceived exponential relationship between correct the relationship between motion blur and correct iris 474
425 recognition rate and sample iris diameter (Figure 3) is not identification. 475
426 intuitive. One would expect that as the number of pixels in 476
an iris image decreases, the correct recognition rate would 5.2. Iris Recognition Software
427 477
decrease linearly. The CRR does not drop below 95% until
428 GIRIST was selected largely because of its performance 478
the sample iris diameter is decreased to below 70 pixels.
429 The next drop off point occurs at sample iris diameter 49. and its availability. While it is not open source, the 479
430 These saturation points are important because increasing executable is free. No other iris recognition software was 480
431 beyond the resolution suggested by those points does not used to corroborate our results. Other iris recognition 481
432 seem to provide a substantial increase in benefit. software may produce different results. 482
433 483
434 4.2. Resolution Impact 5.3. Image Database 484
435 485
Figure 4 presents CRR results for the iris blur A single database was used in this research [15].
436 486
represented by Table 2. Therefore the limits are limited to this single database. For
437 example, the diameters of the irises in these images are 487
438 around 90 pixels. We are limited by this maximum 488
439 resolution. The behavior of CRR for iris diameters above 489
440 90 pixels is unknown. Another example is that the images 490
441 contained within this database are not color. The effect of 491
442 color images is unknown. 492
443 493
444 6. Conclusion & Future Work 494
445 495
446 Iris recognition is increasingly employed as a biometric 496
modality to improve the security across the government,
447 497
public, and private sectors; however the limitations of iris
448 498
recognition are yet to be fully understood. This work
449 explored the impact of image resolution (with
499
corresponding iris width) and image blur on correct
recognition rates.
Figure 4: Correct Recognition Rate as a function of
Gaussian Blur Window Size

4
IJCB IJCB Submission ****. CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE. IJCB
#**** #****

500 [8] J. Daugman, "Iris Recognition and Anti-Spoofing 550


501 Results received as a part of the iris resolution Countermeasures," The 7th International Biometrics 551
investigation indicate the iris may be significantly smaller Conference, 2004.
502 552
than what was previously perceived. The image resolution [9] U. K. A. Nadeau, P. Lopez-Meyer, and S. Schuckers, "Iris
503 Identification from a Distance," Honors Thesis, Clarkson 553
504 analysis saturation points provide evidence that resolution University Department of Electrical and Computer 554
505 may be sacrificed at the expense of perhaps a more Engineering, Clarkson University. 555
506 economical camera, a more portable camera, a more [10] I. O. f. Standardization, "Information Technology - 556
507 convenient camera such as a webcam, or data storage. Biometric data interechange formats " in Part 6: Iris image 557
508 Though saturation points are apparent in the preceding data vol. ISO/IEC 19794-6:2005, ed: International
558
data, it is uncertain whether there are further saturation Organization for Standardization, 2005, p. 25.
509 [11] K. B. K. Hollingsworth, and P. Flynn, "Pupil Dilation 559
points at increasing resolution levels. This necessitates
510 Degrades Iris Biometric Performance," Computer Vision and 560
further research into the relationship between image
511 resolution and correct iris identification. Image Understanding, vol. 113, 2009. 561
512 [12] L. R. X. Huang, and R Yang, "Image Deblurring for Less 562
513 Intrusive Iris Capture," IEEE Conference on Computer 563
The effect of increasing image blur on correct Vision and Pattern Recognition, 2009.
514 recognition is important in that it is now quantitatively 564
[13] R. M. Y. Lee, and P. Phillips, "Robust Iris Recognition
515 defined. It is clear that a small amount of blur can be Baseline for the Ocular Challenge," IEEE Conference on 565
516 employed to reduce potential image noise without losing Automatic Face and Gesture Recognition, 2011. 566
517 much iris recognition accuracy. [14] J. D. a. C. Downing, "Effect of Severe Image Compression 567
518 on Iris Recognition Performance," IEEE TRANSACTIONS 568
519 We feel that the presented results are useful for the ON INFORMATION FORENSICS AND SECURITY, vol. 3, 569
520 practitioners that use iris-identification systems because pp. 52-61, 2008. 570
[15] UBIRIS.v1 [Online]. Available:
521 they a allow better understanding of the limitations and http://iris.di.ubi.pt/ubiris1.html 571
522 capabilities or such systems. [16] H. Davis, Practical Artistry: Light & Exposure for Digital 572
523 Photographers: O'Reilly Media, 2008. 573
524 Additional research should be conducted to understand [17] T. Ang, Dictionary of Photography and Digital Imaging: 574
525 the impact of different types of blur including motion blur, The Essential Reference for the Modern Photographer: 575
526 resolution degradation on a wider range of iris recognition WatsonGuptill, 2002. 576
techniques, and image databases. [18] S. Chernenko. Gaussian Filter, or Gaussian Blur. Available:
527 http://www.librow.com/articles/article-9
577
528 [19] "GRUS IRIS TOOL ", ed. 578
529 7. Conclusion & Future Work [20] P. G. R. M. a. P. J. Phillips, "Face Recognition Vendor Test 579
530 This study is supported by a federal grant. Grant 2002 Performance Metrics," Fourth International 580
531 identification removed for double blind review. Conference on Audio-Visual Based Person Authentication, 581
532 2003. 582
[21] CASIA-IrisV3 [Online]. Available:
533 References http://www.cbsr.ia.ac.cn/english/IrisDatabase.asp
583
534 [1] B. L. Jay R Bhatnagar, RK Patney, "Performance issues in [22] MMU1 Iris [Online]. Available: 584
535 biometric authentication based on information theoretic http://pesona.mmu.edu.my/~ccteo/ 585
536 concepts: A review," IEEE Technical Reviews, vol. 27, pp. [23] GIRIST Tutorial. Available: 586
537 273-285, 2010. http://www.grusoft.com/girist/GIRIST_tutorial.pdf 587
538 [2] A. K. R. Jain, Arun; Prabhakar, Salil, "An introduction to 588
539 biometric recognition," IEEE Transactions on Circuits and 589
Systems for Video Technology, vol. 14, pp. 4-20, 2004.
540 [3] L. F. a. A. Safir, "Iris Recognition System," US Patent 4 641
590
541 649, 1987, 1987. 591
542 [4] J. Daugman, "The Importance of Being Random: Statistical 592
543 Principles of Iris Recognition," Pattern Recognition, vol. 36, 593
544 2003. 594
545 [5] T. T. L. Ma, Y. Wang, and D. Zhang, "Efficient Iris 595
Recognition by Characterizing Key Local Variations," IEEE
546 596
Transactions on Image Processing, vol. 13, 2004.
547 [6] Iris Recognition. Available: 597
548 http://www.biometrics.gov/Documents/irisrec.pdf 598
549 [7] R. K. G. Annapoorani, P. Gifty Jeya, and S. Petchiammal, 599
"Accurate and Fast Iris Segmentation," International Journal
of Engineering Science and Technology, vol. 2, 2010.