High confidence visual recognition of persons by a test of statistical independence
Summary (7 min read)
I. INTRODUCTION
- FFORTS to devise reliable mechanical means for bio-E metric personal identification have a long and colorful history.
- Other biometric identifiers that have been adopted historically, ranging from cranial dimensions to digit length, as well as some of the numerous geometric facial measurements currently being tried, are described in [17] , [25] .
- The present report resolves all of these questions affirmatively and describes a working system.
A. Operators for Locating an Iris
- Iris analysis begins with reliable means for establishing whether an iris is visible in the video image, and then precisely locating its inner and outer boundaries (pupil and limbus).
- The complete operator behaves in effect as a circular edge detector, blurred at a scale set by o, that searches iteratively for a maximum contour integral derivative with increasing radius at successively finer scales of analysis through the three parameter space of center coordinates and radius ( T O , yo, T ) defining the path of contour integration.
- Then in the subsequent interior search for the pupillary boundary, the arc of contour integration ds in operator (1) is restricted to the upper 270" in order to avoid the corneal specular reflection that is usually superimposed in the lower 90" cone of the iris from the illuminator located below the video camera.
- With o automatically tailored to the stage of search for both the pupil and limbus, and by making it correspondingly finer in successive iterations, the operator defined in (1) has proven to be virtually infallible in locating the visible inner and outer annular boundaries of irises.
- Using multigrid search with gradient ascent over the image domain (z,g) for the center coordinates and initial radius of each series of contour integrals, and decimating both the incremental radius interval Ar and the angular sampling interval A8 in successively finer scales of search spanning four octaves, these iris locating operations become very efficient without loss of reliability.
B. Assessing Image Quality, Eyelid Occlusion, and Possibility of Artifice
- The operators previously described for finding an iris also provide a good assessment of "eyeness," and of the autofocus performance of the video camera.
- Excessive eyelid occlusion is alleviated in cooperating Subjects by providing live video feedback through the lens of the video camera into which the Subject's gaze is directed, by means of a miniature liquid-crystal TV monitor displaying the magnified image through a beamsplitter in the optical axis.
- A further test for evidence that a living eye is present exploits the fact that pupillary diameter relative to iris diameter in a normal eye is constantly changing, even under steady illumination [ 11, [ 111.
- Continuous involuntary oscillations in pupil size, termed hippus or pupillary unrest, arise from normal fluctuations in the activities of both the sympathetic and parasympathetic innervation of the iris sphincter muscle [ 11.
C. Two-Dimensional Gabor Filters
- An effective strategy for extracting both coherent and incoherent textural information from images, such as the detailed texture of an iris, is the computation of 2-D Gabor phasor coefficients.
- This family of 2-D filters were originally proposed in 1980 by Daugman [8] as a framework for understanding the orientation-selective and spatial-frequency-selective receptive field properties of neurons in the brain's visual cortex, and as useful operators for practical image analysis problems.
- Their mathematical properties were further elaborated by the author in 1985 [9], who pointed out that such 2-D quadrature phasor filters were conjointly optimal in providing the maximum possible resolution both for information about the orientation and spatial frequency content of local image structure ("what"), simultaneously with information about 2-D position ("where").
- 2-D Gabor functions can form a complete self-similar wavelet expansion basis [lo], with the requirements of orthogonality and strictly compact support [20]-[21] relaxed, by appropriate parameterization for dilation, rotation, and translation.
FREQUENCY RESPONSE
- EQUATION where the substituted variables (2': y') incorporate dilations of the wavelet in size by 2-", translations in position @, q), and rotations through angle 6': EQUATION ).
- It is noteworthy [9] that as consequences of the similarity theorem, shift theorem, and modulation theorem of Fourier analysis, together with the rotation isomorphism of the Fourier transform, all of these effects of the generating function.
D. Doubly Dimensionless Projected Polar Coordinate System
- Zones of analysis are established on the iris in a doubly dimensionless projected polar coordinate system.
- Its purpose is to maintain reference to the same regions of iris tissue regardless both of pupillary constriction and overall iris image size, and hence regardless of distance to the eye and video zoom factor.
- The homogeneous rubber sheet model assigns to each point in the iris, regardless of size and pupillary dilation, a pair of dimensionless real coordinates ( T , e) where T lies on the unit interval [0,1] and 0 is the usual angular quantity that is cyclic over [0,2a] .
- The zones of analysis always exclude a region at the top of the iris where partial occlusion by the upper eyelid is common, and a 45" notch at the bottom where there is a corneal specular reflection from the filtered light source that illuminates the eye from below.
- Rotation invariance to correct for head tilt and cyclovergence of the eye within its orbit is achieved in a subsequent stage of analysis of the iris code itself.
111. CODE CONSTRUCI'ION AND ENTROPY MEASURES
- An uncompressed code length of 256 bytes was chosen because this is roughly the capacity of the three-channel magnetic stripe affixed to the reverse side of the standard IS-7811 credit/debit card [3].
- But this absolute code length only establishes an upper bound on the information capacity of an iris code, and it is important to know its actual inherent capacity.
- It is then also important to know the "source entropy" associated with the typical human iris signal, which will be much less than the upper bound determined by the resolution of imaging, because of inherent correlations (especially radial) within the iris.
- These reduced entropies directly influence the confidence levels associated with any decision strategy.
- Such a test of statistical independence is passed almost certainly for two iris codes from different eyes, but the same test is failed almost certainly when the compared signatures originate from the same eye.
B. Commensurability of Iris Codes
- A critical feature of this coding approach is the achievement of commensurability among iris codes, by mapping all irises into a representation having universal format and constant length, regardless of the apparent amount of iris detail.
- It is not obvious mathematically how one would make objective decisions and compute confidence levels on a rigorous basis in such a situation.
- This difficulty has hampered efforts to automate reliably the recognition of fingerprints.
- Commensurability facilitates and objectifies the code comparison process, as well as the computation of confidence levels for each decision.
- It thereby greatly increases both the speed and the reliability of iris recognition decisions.
D. Number of Independent Degrees-of-Freedom in an Iris Code
- Such a code possesses far fewer than 2 048 independent binary degrees-of-freedom.
- A given furrow or ciliary process tends to propagate across a significant radial distance in the iris, exerting its influence on several remote parts of the code, thus reducing their independence.
- Finally, inherent correlations are introduced by the bandpass property of the 2-D Gabor filters, specifically by the finite bandwidth determined by parameters a, p, and w in (13).
- Even though the peak response of the bandpass filter might be at a very high frequency, its passband introduces phase coherence that lingers for a greater number of cycles, the narrower its bandwidth.
- The number of independent degrees-of-freedom typically remaining in an iris code after both of these sources of correlation have been factored in (those arising from the 2-D Gabor filters and those inherent within an iris), can be estimated by examining the distribution of Hamming distances computed across a population of unrelated iris codes.
Hamming Distances for Imposters
- A theoretical plot of the probability density function associated with such a binomial process having N = 173 and p = 0.5 is also shown in Fig. 6 as a smooth curve, and it offers a good fit to the data.
- In summary it appears that there exist the equivalent of about 173 independent binary degrees-offreedom typically remaining in a 2 048-bit iris code, once both the correlations introduced by the 2-D Gabor filters and those inherent in the iris have been factored in.
- The problem of recognizing the signature of a given iris as belonging to a particular individual, either after exhaustive search through a large database or just by comparison with a single authentication template, can be formulated within the framework of statistical decision theory [22], [27].
- This framework also resolves the critical problem of assigning a confidence level to any such recognition decision.
- By this approach the authors can convert the problem of pattern recognition into a much more expedient task, which is the execution of a.
Solid curve is (22).
- A. Nevman-Pearson Formalism of bits from different iris codes disagree.
- The actual distribution of observed Hamming distances between codes for different irises is shown in Fig. 6 , which is generated from 2 064 complete comparisons between unrelated pairs of iris codes.
- In the present application the four possible outcomes are termed Acceptance of Authentic (AA), Acceptance of Imposter (IA), Rejection of Authentic (AR), and Rejection of Imposter (IR).
- The goal of the decision-making algorithm is to maximize the conditional probabilities of AA and IR, while minimizing the likelihoods of IA and AR.
B. Strategies and Decidability
- It is clear that the four probabilities separate into two pairs that must sum to unity, and two pairs are governed by inequalities: would be reduced if their two means were farther apart, or if their variances were smaller, of both.
- Of course, the two distributions in general will not be matched in form and variance, as was implied in Figure 7 for simplicity.
- Such a decision strategy diagram, sometimes called a receiver operating characteristic or Neyman-Pearson curve, plots P(AA) from (23) against P(IA) from (25) as a locus of points.
- Clearly, strategies that were excessively conservative or excessively liberal would correspond to sliding along the curve towards the two diagonal extremes.
- This distance is monotonically related to the quantity d', for "detectability" or "decidability," defined as the difference between the means of the two distributions that were shown schematically in Fig. 7 divided by a conjoint measure of their standard deviations.
A. Database
- The performance results reported here are based partly on a photographic database of eye images generously made available in 1989 by Ophthalmology Associates of Connecticut, which were digitized and then combined with further databases of images subsequently acquired directly with video cameras in Massachusetts and in Cambridgeshire, England.
- Multiple images were always acquired from each person, ranging from 2 to 10 images of each eye over the time period (average 3.04 images per eye).
- Images in RS-170, VHS (NTSC), and S-VHS (NTSC) formats were digitized by 480 x 640 monochrome 8-bit/pixel framegrabber boards in either Macintosh or (by SCSI interface) SUN sparcstation hosts.
- Image resolution and iris size within the images varied due to both distance and video zoom factor, but the outer diameter of the iris was always greater than 60 pixels and was usually in the range of 100 pixels to 200 pixels.
- Ethnic groups and nationalities represented in the combined databases included persons of Northern European, Mediterranean, Eastern European, Indian, Semitic, Afro-American, Hispanic-American, Japanese, and Chinese origin.
B. Imposters ' Humming Distances
- The distribution of Hamming distances generated by 2 064 direct comparisons between painvise unrelated iris codes was seen previously in Fig. 6 .
- The raw distribution was well described by a suitably fitted binomial model, whose effective number of implicit Bernoulli trials was appropriately reduced to factor out the residual correlations that exist among the bits within a given iris code.
- Because of possible cyclovergence of the eye in its orbit as well as tilting of the head, all iris code comparisons must be performed over a range of relative orientations.
- The comparison process then becomes a "best of 71' ' test of agreement, and this must be factored into the statistical decision theory that underlies this method of personal identification.
- The distribution is biased toward a lower mean Hamming distance of 11 = 0.450, since only the best level of agreement after all seven rotations (i.e., the smallest Hamming distance) is kept and registered as the degree of match.
D. Equivalent Bernoulli Trials
- The distributions of Hamming distances for 2,064 pairwise comparisons of "imposters" (summed across pairwise unrelated iris codes), and for 1208 pairwise comparisons of "authentics" accumulated separately, are shown together for comparison in Fig. 11 .
- They are clearly well separated, with no empirical overlap and with no observations whatever falling in the region of 0.25 to 0.35 Hamming distance.
- The authors have seen that on average, when comparing two iris codes obtained at different times from the same ("authentic") iris and making provision for possible headleye tilt, any pair of corresponding bits have a probability of 0.084 of not matching.
- Needless to say, sufficiently many tosses could resolve the question about which type of coin it was with enormously high confidence.
- The shapes of the two distributions shown in Figure 11 would have been expected using about 480 tosses of the p = 0.450 coin, and using about 40 tosses of the p = 0.084 coin, respectively, in each run of trials.
E. Decision Confidence Levels
- The Bernoulli representation noted above for this pattern recognition task clarifies the calculation of confidence levels associated with any decision, including extrapolation of confidence levels into the region between the two distributions where no Hamming distances were observed empirically.
- As specified in (23)-( 26), the conditional probabilities of personal identity or nonidentity given a particular observation can be calculated as the cumulative integrals under the two density distributions, taken from opposite directions up to whatever Hamming distance was observed.
- Empirically, comparisons of iris codes computed from the available database of eye images produced no Hamming distances in the range of 0.25 to 0.35, so the use of any criterion in this range would produce 100% correct performance.
- The means of the two distributions in Fig. 11 indicate typicality.
- In the typical authentic comparison, which generates a Hamming distance of only 0.084, the confidence with which the Subject is accepted (given this observation) corresponds to a conditional false accept probability of one in.
F. Ergonomics, Robustness to Noise, and Imaging Factors
- In many respects, the iris of the eye is inherently difficult to image at a comfortable "social" distance (e.g., several feet from a mounted video camera).
- More critical even than these limitations of spatial resolution is the limitation of greyscale resolution, since without appropriate gain control of the video signal, many very darkly pigmented irises tend to be digitized flatly into only the lowest few states of an 8-bit A-to-D converter and thus reveal little structure.
- All of these factors contribute to the observation that different images of the same eye at different times may generate iris codes that disagree in as many as 25% of their bits (the highest observed Hamming distance in Fig. 10 , for "authentics").
- Thus, the hypothesis of independence can be strongly rejected over all but a narrow range of possible Hamming distances.
- It is perhaps illuminating that at the "cross-over''.
G. Speed of Decision Making
- The Bernoulli trial XOR formulation of the decision problem allows us to exploit the 32-bit architecture of a CPU for 16-fold parallelization.
- As a result, on a RISC general-purpose CPU any "presenting" iris code can be compared exhaustively against a large database of stored codes in search of a match at the rate of about 4 000 per second.
- The simple 74F86 integrated circuit contains four independent XOR gates that can be clocked at 80 megahertz.
- Rather, here he only needs to present his eye to the camera, and his identity is rapidly and automatically determined without any further interaction, by exhaustive search through a database that might be extremely large.
- As Shakespeare conveyed it much less mechanically in The Merchant of Venice (Act I, Scene l), in the tradition of conceiving the eyes as windows to the soul, "Sometimes from her eyes I did receive fair speechless messages.".
VI. CONCLUSION
- Aristotelian philosophy held that the ELSO< (%dos, distinguishing essence) of something resided in that quality which made it different from everything else.
- When the authors need to know with certainty who an individual is, or whether he is who he claims to be, they normally rely either upon something that he uniquely possesses (such as a key or a card), something that he uniquely knows (such as a password or PIN), or a unique biological characteristic (such as his appearance).
- Technologically the first two of these criteria have been the easiest to confirm automatically, but they are also the least reliable, since (in Aristotelian terms) they do not necessarily make this individual different from all others.
- Today, the authors hold that the uniqueness of a person arises from the trio of his genetic genotype, its expression as phenotype, and the sum of his experiences.
- It is hard to imagine one better suited than a protected, immutable, internal organ of the eye, that is readily visible externally and that reveals random morphogenesis of high statistical complexity.
Did you find this useful? Give us your feedback
Citations
4,017 citations
2,829 citations
Cites methods from "High confidence visual recognition ..."
...Altogether 2,048 such phase bits (256 bytes) are computed for each iris, but in a major improvement over the earlier ( Daugman 1993 ) algorithms, now an equal number of masking bits are also computed to signify whether any iris region is obscured by eyelids, contains any eyelash occlusions, specular reections, boundary artifacts of hard contact lenses, or poor signal-to-noise ratio and thus should be ignored in the demodulation code as ......
[...]
...Algorithms described in ( Daugman 1993, 1994 ) for encoding and recognizing iris patterns have been the executable software used in all iris recognition systems so far deployed commercially or in tests, including those by British Telecom, US Sandia Labs, UK National Physical Lab, NBTC, Panasonic, LG, Oki, EyeTicket, IBM SchipholGroup, Joh.Enschede, IriScan, Iridian, and Sensar....
[...]
2,437 citations
Cites methods from "High confidence visual recognition ..."
...Only phase information is used for recognizing irises because amplitude information is not very discriminating, and it depends upon extraneous factors such as imaging contrast, illumination, and camera gain....
[...]
...The author’s algorithms [8]–[10] for encoding and recognizing iris patterns have been the executable software used in all iris recognition systems so far deployed commercially or in tests, including those by British Telecom, Sandia Labs, U.K. National Physical Lab, Panasonic, LG, Oki, EyeTicket, Sensar, Sarnoff, IBM, SchipholGroup, Siemens, Sagem, IriScan, and Iridian....
[...]
...…have been the executable software used in all iris recognition systems so far deployed commercially or in tests, including those by British Telecom, Sandia Labs, U.K. National Physical Lab, Panasonic, LG, Oki, EyeTicket, Sensar, Sarnoff, IBM, SchipholGroup, Siemens, Sagem, IriScan, and Iridian....
[...]
2,046 citations
Cites background from "High confidence visual recognition ..."
...It is a thin diaphragm stretching across the anterior portion of the eye and supported by the lens (see Fig....
[...]
...Last, Section IV provides concluding observations....
[...]
...Unlike the human face, however, the variability in appearance of any one iris might be well enough constrained to make possible an automated recognition system based on currently available machine vision technologies....
[...]
...…data base entry; 2) choosing a representation of the aligned iris patterns that makes their distinctive patterns apparent; 3) evaluating the goodness of match between the newly acquired and data base representations; 4) deciding if the newly acquired data and the data base entry were derived…...
[...]
1,852 citations
References
65,425 citations
20,442 citations
20,028 citations
15,525 citations