KR100307792B1 - Biometric measurement based on iris analysis - Google Patents

Abstract

The image analysis algorithm finds the iris in a live video image of a person's face (10) and encodes its texture as an "iris code" (24). Iris textures are extracted at multiple analytical scales by a set of self-like quadrant bandpass filters defined in a dimensionless polar coordinate system. The sign of the projection to these filters of many different parts of the iris determines each bit in the iris code. The comparison between the codes is done immediately by an exclusive OR (XOR) logical operation. Pattern recognition is achieved by combining signal processing methods with statistical decision theory and leads to a statistical check of independence based on a similarity measure (hamming distance) calculated from the XOR of any two iris codes (26). This measurement establishes, verifies, or denies the identity of any individual. It also generates an objective level of confidence associated with the identification decision (30).

Description

[Name of invention]

Biometric person identification method based on iris analysis

[Background of invention]

FIELD OF THE INVENTION The present invention relates to personal identification, and in particular relates to the field of automated human identification by biomeric indicia.

Identification is as old as humanity itself. As technology and services develop in the modern world, human activities and transactions thrive, requiring rapid and reliable personal identification. Examples include, but are not limited to, passport control, computer login control, bank automatic teller machine and other transaction confirmations, premises access control and security systems. All of these identification efforts share a common goal of speed, reliability and automation.

The use of biometric indicators for identification should be unique, uniquely measurable and unique over time. While many indicators have been proposed over many years, fingerprints are probably the most common form of successful biometric identification system. As is known, the two fingerprints are not identical to each other, and they do not change except in the case of injury or surgery. However, identification through fingerprints requires physical contact with the person. It is obvious that there are obvious drawbacks. There is no way to get fingerprints from a distance, and there is no such way.

An iris is a biometric indicator that has been largely overlooked in the art. The iris of everybody's eye has a very complex inherent texture and is known to remain essentially unchanged during its survival. Even if they are the same, no two irises are identical in texture or detail. As the internal organs of the eye, the iris is well protected from the outside environment, but can be easily seen a few yards away from the colored disks surrounded by the white tissue of the eye behind the clear protective window of the cornea of the eye. The iris contracts or relaxes to resize the pupil in response to debt, but its fine texture does not change significantly apart from contraction and stretch. Deformations in such textures can be instantly converted mathematically in analyzing the iris phase and extract and encode iris signatures that remain the same over a wide range of pupil dilation. The richness, uniqueness and invariability of iris textures as well as external visibility make the iris suitable for highly reliable automated personal identification. Registration and identification of the iris can be done automatically and unobtrusively using a video camera without any physical contact.

By comparison, other biometric measurements, such as signs, photographs, fingerprints, voice prints, and retinal vascular patterns, all have significant defects. Although signs and photos are inexpensive and easy to obtain and store, they cannot be identified automatically with confidence and can be easily forged. Electronically written gates are susceptible to fluctuations in the human voice, and they can also be forged. Fingerprints or hand prints require physical contact, and they can also be forged and artificially damaged.

Iris identification is not to be confused with retinal identification. The iris is easy to see and can be shot immediately by a video camera. On the contrary, the retina is hidden deep in the eye and difficult to see. Common conditions, such as small pupils or cataracts, make it difficult or impossible to see the retina, but do not affect the visibility of the iris.

US Pat. No. 4,641,349, hereinafter referred to as '349,' patented to Flom and Safir under the name "Iris Recognition System", the only previous attempt to use this good feature of the iris in a personal identification system ). This' 349 reference discloses a general concept of using the iris as an identification method but does not describe an advanced embodiment of such a system. It also does not disclose automatic means for finding and isolating the iris in the image regardless of the size and position of the iris and extracting and encoding its texture. In addition, the '349 reference does not describe any method of estimating identification decisions when a series of features have been edited. In general, the feature lists from the two iris images will be partially coincident and partially inconsistent whether they were created from the same iris or not. In general, such lists also differ in the number of features that make up them. No theoretical or mathematical formula is provided for making decisions based on a comparison between such discordant data sets. In addition, no method of calculating the confidence level associated with identification is disclosed.

[Summary of invention]

It is a broad object of the present invention to provide a system for identifying a person based on the iris of one eye.

Another object of the present invention is to provide a system for identifying a person very quickly and reliably in about 1 second.

It is another object of the present invention to provide an identification system that calculates a confidence level for any identification decision based on objective and stringent criteria.

It is yet another object of the present invention to provide an identification system that can be identified without the operation of the subject and without physical contact with the subject.

It is yet another object of the present invention to provide an identification system that can distinguish between live subjects and imposters using dead replicated means of identification.

These and other objects are achieved in the present invention by a method of uniquely identifying a particular person having the following steps. First, the system acquires a digitalized image of the eye of the person who is trying to identify through a video camera. Then, if there is an iris in the image, separate it to define a circular pupil boundary between the iris portion and the pupil portion of the image, and use an arc that is not concentric with the pupil boundary. sclera) defines another circular boundary between parts. The system of the present invention sets up the polar coordinate system on the separated iris so that the origin of the coordinate system is the center of the circular pupil boundary. Next, a plurality of annular analysis bands are defined within the iris phase, which do not include any preselected portions of the iris that can be blocked by mirror reflection from the eyelids, eyelashes or illuminator. do.

The iris phase portion within these annular analysis bands is analyzed and encoded using special signal processing means with a multi-scale, self-similar set of quadrature bandpass filters at polar coordinates. This results in an iris code having a fixed length for all irises and having a universal format. The resulting code is stored as a reference code. Because of the general format and length of all such iris codes, the comparison between different iris codes is very efficient and simple. In particular, the comparison between any two iris codes is accomplished by calculating a basic logical XOR (exclusive-OR logical operation) between all the bits corresponding to them, and then calculating the norms of the resulting binary vector. This comparison measure may also be described as the Hamming distance of two iris code vectors. The universal format of the iris code also helps to quickly parallel search a large database of stored reference iris codes to determine an individual's identity.

In order to perform identification using the reference code, the system generates an identification code from the identification subject according to the providing step, the defining step, the determining step, the establishing step, the description step, and the analysis step. The system then compares the identification code with the reference code to determine the hamming distance between these codes. This distance is then converted to the probability calculated from the same iris, ie the same person, by calculating the probability that the observed fragments of bits in the two codes can coincide if the two codes are independent of each other. . The predetermined criterion applied to this measured hamming distance generates a "yes" or "no" decision, and the confidence level for that decision is provided by the calculated probability.

Other and further objects of the invention, along with various novelty features which characterize the invention, are pointed out with specificity in the claims appended hereto to form part of the disclosure. In order to better understand the present invention, its operational advantages and the specific objects obtained by using the same, reference should be made to the accompanying drawings and the illustrative examples which describe preferred embodiments of the present invention.

[Brief Description of Drawings]

1 is a block diagram illustrating the main steps of the process of the present invention.

2 is a view showing photographs of human eyes superimposed to illustrate the process of the present invention.

3 is a diagram illustrating a quadrature band pass filter used by the present invention as an image convolution kernel to extract iris structures at many analysis scales.

4A to 4C show a series of plots illustrating how iris image data is converted into iris code bits by a four-minute bandpass filter in accordance with the present invention.

5 shows a plot showing the Hamming distance for "fake", ie for comparison between iris codes calculated from different irises.

6 shows a plot showing the formula for statistical decision theory.

7 shows a plot illustrating the theoretical effect of the treatment of decision criteria.

FIG. 8 is a plot showing the Hamming distance for "authentics", ie comparing plots between iris codes computed from the same iris in different cases.

9 shows a plot showing Hamming distances for the combined real and fake.

10 shows a plot showing a binomial distribution suitable for a Hamming distance distribution.

11 shows a plot showing an error as a function of a Hamming distance criterion.

12 shows a table of performance rates achieved with the present invention.

Detailed Description of the Invention

One embodiment of the present invention is shown in schematic form in FIG. 1 and consists of a block diagram showing the important steps in creating an iris identification code for a person and then using the code to perform identification. The process is discussed in the general section, followed by a detailed analysis.

The iris of the human eye is a complex structure consisting of muscles, trabecular meshwork of pectinate ligaments, connective tissue, blood vessels and chromatophores. It is. Externally, contraction furrows, collagenous stromal fibers, filaments, serpentine vasculature, rings, follicles, hypothalamus Represents a visible test with both radial and angular variations from freckles: all of which together form a unique "fingerprint". This magnified optical image of the human iris constitutes a plausible biometric signature for establishing or confirming personal identification. Other characteristics of the iris that serve this purpose and potentially have superior status over fingerprints for automatic identification systems include the inability to surgically change its texture without taking great risk; Intrinsic protection and isolation from the physical environment are the internal organs of the eye behind the cornea and aqueous humor, and its easily monitored physiological response to light provides a natural test for the artefact. Other technical advantages over fingerprints for automatic recognition systems are the easy registration of the iris optically without physical contact and the inherent polar geometry of the iris that conveys the natural coordinate system and origin.

Until studying the present invention, it is not known whether there is sufficient freedom or variation in the iris depending on the individual to give it specificity as a conventional fingerprint. In addition, the detailed iris structure can be reliably extracted from the video image (of muscle length compared to normal data size) to generate compact code and to determine identity with high statistical confidence in less than one second on conventional devices. It was also opaque whether enough algorithms could be developed to do it completely within a small processing time. The present invention clearly solves all these problems.

The system of the present invention can be viewed in five stages. First, the image of the eye to be analyzed must be obtained in a digital form suitable for analysis as shown in block 10 of FIG. Next, the iris portion on it must be defined and separated (blocks 12, 14, 16 and 18). The defined region of the phase is then analyzed to generate the iris code (block 20). It is noted that the first iris code generated for a particular iris is stored as a reference code (block 22). The system then uses the reference code to perform identification by comparing the reference code with the provided code (block 24) to obtain a hamming distance (step 26). This data allows the system to establish, confirm or disconfirm the subject's identity (block 28) and to calculate a confidence level for the decision (block 30). The speed of this process also allows for manual identification by exhaustive search of many databases, rather than just identifying a single authorization code.

In practical application of the present system, the decitalized image as shown in FIG. 2 shows an eye with an iris 102 surrounding the pupil 104. The sclera portion 105 of this eye, which becomes a white portion, surrounds the iris 102. The first step of processing the image locates the pupil boundary 106 and separates the pupil 104 from the iris 102 with a high degree of accuracy (FIG. 1, block 12). This step is very important to ensure that the same part of the iris is assigned the same coordinate each time the image is analyzed, regardless of the degree of pupil dilation.

The inner boundaries of the iris forming the pupil can be accurately determined using the fact that the boundaries of the pupil are essentially circular edges. As can be seen in FIG. 2, the pupil 104 is generally dark while the iris 102 is brighter and variously colored. However, this relationship may sometimes be reversed, for example in the case of an eye with a dark iris and a slightly turbid inner lens, or in the case of optically coaxial illumination (directly to the eye), in which case the light is returned from the retina. Reflects and exits through the pupil. Another reason why the pupil images may be bright is due to mirror reflections from the cornea. The universal method of finding the pupil boundary must be certain enough to operate reliably even if the area of the pupil is actually darker than the iris or not. In the present invention, a system has been developed which combines evidence for actual pupil boundaries with the desired certainty of motion and accuracy.

The method of the present invention detects the pupil boundary as a sudden change in brightness when summed along a circle with a constantly increasing radius. This sudden change is maximal when the circle has its center near the actual pupil center and when its radius is equal to the actual pupil radius. As such, the image processing problem of finding a pupil can be formulated as an optimization problem, in which case a series of "exploding circles" have a number of starting points on the grid whose center coordinates are constant. (trial point) to be positioned at each of them. For each expanding circle, and for each value of its radius, the overall image brightness is summed for a number of points placed on this circle (using a certain number, typically 128 points on each circle, is simply the circumference Avoid an automatic increase in the sum of the brightnesses due to the increase). The system retrieves the maximum rate of change in this amount as the radius expands. For a coordinate source that best describes the pupil boundary, there will be a sudden "spike" in the rate of change of the sum of the illuminations around that boundary when the radius matches the radius of the pupil relay. This spike will be larger for a circle that shares the center coordinates and radius of the pupil than for all other circles. In this way, the problem of precisely positioning the pupil has been translated into an optimal problem, in which case the 3-variable space retrieves the best combination of circle center coordinates (x ₀ , y ₀ ) and radius r.

The process of the invention sums up the contour integral of I (x, y), which is the phase intensity on the arc ds of a circle with radius r and center coordinates (x ₀ , y ₀ ), It can then be described mathematically by calculating this positive one-way function for r as the radius increases. The maximum absolute value of this derivative is searched for the space of three variables (x ₀ , y ₀ , r):

The partial derivative function for r can be a smoothed or blurred derivative of noise immunity, and it also translates into a percentage change for improved noise immunity (current value of round integrals). Can be divided into

This method also has some inherent noise immunity since the perturbation integration essentially integrates the data according to the perversion, and therefore tends to average the anomalous deviation in pixel brightness.

The search process in the 3-variable space is dominated by gradient ascent, or "hill-climbing". If the candidate series of expansion circles is partially within the pupil, the positive value defined in equation (1) will be larger than for other circles that are not. The closer the center of the concentric circles is to the actual center of pupil, the larger this amount will be. Likewise, the amount in equation (1) will be greater for a circle with a suitable radius. That is, the method can find the optimal combination of three variables by the iterative search process, in which the step size of the change of the three variables is reduced with each successive iteration. The method converges quickly by always moving in the best direction in the 3-variable space (maximum improvement) to reduce these step sizes in each successive iteration. Typically after only four or five iterations, the optimal values of the three variables are determined within a range of less than one pixel. The values for (x ₀ , y ₀ , r) determine the estimated pupil boundary as well as the origin of the polar coordinate system for subsequent iris analysis.

In addition, this effective method of finding and tracking pupil boundaries provides an important safeguard against fakes. One sure way to disable the identification system based on the iris pattern is to provide a video camera with a picture of another person's eyes or even wear a contact lens printed with a certified iris. An important feature of living eyes, however, is that they continue to undergo small vibrations ("hippus") once or twice per second, even under uniform illumination. Contact lenses printed on photographs or images of the iris do not exhibit this variation over time. Since the above process for finding and tracking pupil boundaries is very fast, it is possible to continuously acquire several images and monitor the pupil diameter over time. The absence of hippus oscillations or other small variations in the iris pattern over time is indicative of evidence that a photograph or simulacrum has been provided rather than a living iris. The ability to distinguish living irises from counterfeits or photographs is an important security asset made possible by high-speed means of defining and tracking pupil boundaries.

Once the boundary and center have been determined, the next step is to locate the outer boundary of the iris, or limbus, that will meet the sclera. An important point to consider here is that the pupil is not always centered within the iris. Radius lengths to the right and left limbs can vary by as much as 20%, so these two distances must be calculated to produce the proper iris coordinate system. Another consideration is the fact that these areas should be excluded from the iris analysis because the upper and lower eyelids generally blur the upper and lower boundaries of the iris.

A common method, such as an "extension circle" to accurately determine the pupil boundary, can be used to find the outer boundary of the iris through two modifications. First, given the upper and lower eyelid blockages and generally unequal left and right limbic distances, the method is one at 3 o'clock and one at 9 o'clock, respectively, along the horizontal meridian. It is limited to only two arcs with an arc angle of / 4 radians (45 degrees). The distances between these two boundaries on both sides of the iris are measured separately. Second, because of the possible concentric textures in the iris that can produce the maximum in Eq. (1), the circular integral previously used for the pupil boundary is replaced by the area fraction that removes the details of the iris in retrieving the limbus. . In fact, the "expansion circle" is replaced by two horizontal "exploding pie wedges" which search for the maintained illumination stage, meaning the sclera on both sides. As before, the search process maintains one of finding the maximum value of the rate of change of integrated illumination as the radius of expansion increases. After compensating for the increasing area of illumination integration, the maximum in this derivative for the radius remains unchanged to coincide with the exact left and right boundaries of the iris.

Mathematically, this operation is implemented by searching for the value of r (distance from the center of the pupil to the left or to the right) that maximizes the following equation:

Where r ₀ is the pupil radius (previously calculated) and δ is a thin radial shell; In general, 0.1r ₀ ), I (ρ, θ) is the phase intensity, now represented by polar coordinates ρ and θ, Is 0 or the corresponding meridian at 3 o'clock or 9 o'clock, respectively Same as Search for the outer boundary of the iris by calculating this expression for r values between 1.5r ₀ and 10r ₀ (ie 1.5 times to 10 times the pupil radius) to cover the widest possible diameter of the pupil and iris. It was known that it is possible. Similarly, +/- as the arc angle of the integral in equation (2). Choosing / 8 has proven to be a useful horizontal angular delimiter for pie wedges to avoid the upper and lower eyelids. The results calculated from equation (2) are shown in FIG. 2 as a series of white dots 110-1 and 110-r on iris 102, which exactly coincide with the left and right boundaries of the iris.

In summary, equation (1) finds the inner boundary of the iris, that is, the pupil boundary. This equation generates a series of "extended circles" at various center positions, and iteratively combines one combination of the variable (x ₀ and y ₀ center, radius r) with the largest change in the illumination along the circle. Search. Therefore, we find the maximum absolute value of the partial derivative function for r of the circular integration of the illumination along the circle. This search includes x ₀ , y ₀ , and r variables in a highly efficient iteration of gradient-ascent. Equation (2) finds the outer boundary of the iris, that is, the limbus, where the white sclera begins. The same process works as with the expanding circles, as in Eq. (1), but except for (i) blocking the upper and lower eyelids, which can cause difficulties, and (ii) the fact that the iris is less uniform than the pupil, It may have a large "circulat edge" itself that can be awkward. Thus, Equation (2) instead defines a series of "extended pie wedges" at the horizontal meridian (and thus avoids the upper and lower eyelids), integrating illumination within the pie wedge rather than just along a circle. . Therefore, Eq. (2) prescribes the integral of the area in polar coordinates that is differentiated with respect to the radius, rather than the circumferential integration as in Eq. (1).

When the position of the pupil boundary and rim number is established and the origin of the polar coordinate is fixed at the center of the pupil, a series of analysis regions are assigned to the regions of the iris. These are intensively defined in constant straight segments of the radial distance between the pupil and the limbus to achieve size invariance in the code, whatever the overall size of the iris in a given phase. Thus the polar coordinate system for the iris is dimensionless in both its angular coordinates and the radial coordinates. Since the iris can be modeled approximately as a rubber sheet that relaxes and contracts with a pupillary reflex, its texture and markings thus relax and contract. These distortions are eliminated by using radial coordinates that represent the distance as just another part of the total distance from the inner boundary (the pupil) of the iris to its outer boundary. As such, the irises given to different states of pupil dilation should generate approximately the same iris code. Another aim of this dimensionless coordinate system is to ensure that differences in the overall size of the iris image itself, due to strokes at different distances, do not alter the calculated iris code.

Since the pupils are generally not completely horizontally centered in the iris, the fractionation needs to be based on the right and left rim booth values weighted cosineally by the angle. These areas are excluded from analysis and encoding due to frequent partial blockage of the upper iris by the upper eyelid and mirror reflections from the cornea that blur a portion of the lower iris. A description of their ultimate analysis area superimposed on a particular iris image can be seen in FIG.

In particular, the iris portion to be analyzed is mapped and segmented into analysis band 112 (see also FIG. 2). These bands of analysis are defined in a particular polar coordinate system with a radial punctuation that, as is common, may be slightly deformed if the inner and outer boundaries of a particular iris are not concentric. In particular, for each coordinate near the iris, the radial coordinate r of a point is defined by a portion of the distance from the pupil boundary to the sclera along its ray. Thus, just as each coordinate is (traditional) a dimensionless amount between 0 and 360 degrees, the radial coordinate is dimensionless in this bounded system and is independent of the overall size of the iris and the degree of pupil dilation. Regardless, it will always be in the range between 0 and 1. Thus, such a double-dimensionless polar coordinate system is essentially size-invariant (ie, essentially compensates for the variation in distance from the eye to the video camera). Likewise, this coordinate system also essentially compensates for the non-concentricity of the inner and outer boundaries of the iris.

Four additional special features of the analysis band 112 are needed to compensate for the deviation of many iris images from the ideal illusion stereotype. First, because the pupil 104 itself often has irregular boundaries, the innermost analysis zone begins with a radius approximately 1.1 times the average radius of the pupil to completely exclude the pupil. Likewise, since the transition from iris 102 to sclera 105 may be equally irregular and non-circular, the outermost band of analysis is the distance to the outer boundary of the iris (cosineically at intermediate angles). Weighted up to about 80% of the right and left). Third, prepare for blocking the upper and lower parts of the iris by the eyelids; and fourth, prepare for mirror reflections that may cover a portion of the iris when dark lighting is used (usually from below). do. These hypothetical features are excluded by confining the outermost analysis bands to two sectors centered on the horizontal meridian, ie by the upper and lower eyelids and near the 6 o'clock direction of the mirror reflection of the illumination from below. By eliminating a narrow notch it avoids areas that might be blocked. These excluded areas are labeled 114a and 114b in FIG. It is better to divide the iris area into eight annular analysis bands 112 at a portion of the constant radial distance between the inner and outer boundaries defined above.

After accurately defining the phase region to be analyzed, the system processes the data obtained from that region to generate an identification code as shown in block 20 of FIG. Unlike the system described in the prior art, the present invention does not rely on controlling the amount of pupil dilation. Rather, because of the dimensionless radial coordinates that simply measure and divide some of the distance from the inner boundary of the iris to the outer boundary, some of the iris tissues are always within the same band of analysis and how the iris is due to pupil dilation. They will have the same position coordinates regardless of whether they are tensioned or retracted. This dimensionless coordinate system takes advantage of the fact that the relaxation of the iris can be approximated by the relaxation of the rubber plate, and thus his marking can be mathematically restored to an undeformed form because the coordinate system relaxes by the same amount. Therefore, the texture of the iris is always encoded with essentially the same iris code regardless of the degree of pupil dilation and regardless of the overall size of the iris image.

An effective strategy for extracting texture information from phases such as detailed patterns of the iris is convolution with a four-minute bandpass filter such as a 2-D Gabor filter. These 2-D filters are orientation- and frequency-selective receptive observed by inventors in the neurons of the primate visual cortex of primate animals in 1980 and 1985. As a basic framework for understanding field properties, it is also proposed as a useful operator for actual phase analysis problems. Daugman, J. (1980) "Two-Dimensional Spectral Analysis Of Cortical Receptive Field Profiles", Vision Research 20, pp. 847-856 and Dougman, J. (1985) "Uncertainty Relation For Resolution In Space, optimized by space, spatial frequency and orientation optimized by a two-dimensional visual cortical filter. Spatial Frequency, And Orientation Optimized By Two-Dimensional Visual Cortical Filters), "Journal Of The Optical Society Of America, Volume 2 (7), pp. See 1160-1169. As a conjointly optimal filter they provide maximum resolution simultaneously for spatial frequency and orientation information with 2-D position; They inherently achieve a lower bound on binding uncertainty for these four variables, which are governed by the inevitable uncertainty principle. These features are particularly useful for texture analysis because of the location dependence of the texture and the 2-D spectral specificity.

Two of the 2-D Gabor filter families are shown in FIG. 3 with even-symmetric and odd-symmetric waveform profiles along with their contour plots. Defined in many different sizes and positions, these localized, undulating 2-D functions are multiplied by the original image pixel data and integrated in their support area to produce coefficients that describe, extract, and encode image texture information. Applicants named them "2-D Gabor filters" because they were a 2-D generalization of the class of basic functions discussed in 1946 in one dimension by Dennis Garbor. Garbor, D. (1946) "Theory Of Communication" J. Ins. Elec. Eng. Vol. 93, pp. See 429-457.

The 2-D Gabor filter used in the present invention is defined in polar coordinates as follows.

Where r is the radius, Is the angle in radians, Is each frequency And Is a constant.

The real and imaginary parts of the quadrant (preferential and expected) filter pairs from the analytical function are used. Free variables And Is They change together in inverse relation to generate a multi-scale self-similar series of frequency-selective quadrature filters. They are in quadrant because both quadratures are used at each location. They become self-similar because the inverse relationship of all of their magnitude and frequency parameters makes them an extension of each other sharing a common shape. Their positions, defined as ₀ and r ₀ , span the analysis band of the iris.

The manner in which the iris code is generated by passing the 2-D Gabor filter in polar coordinates is shown in Figures 4a, 4b and 4c. The upper trace (figure 4a) represents a 1-D scan around the iris with a certain radius and plots the brightness of the image as a function of the angular coordinates around the iris (for simplicity, this image is shown here As a 1-D signal rather than a 2-D signal). The second trace (Figure 4b) shows the response of the Gabor filter with a particular size and symmetry located at each coordinate corresponding to each of the irises. Note that because of the bandpass nature of Gabor filters, their response to the original input signal can be positive or negative and centered around zero. Slow, non-informative transitions in the brightness of the raw signal, which gradually change up and down around the iris caused by illumination from below, are removed by the bandpass Gabor filter as high frequency noise.

Each bit in the iris code is determined by the positive or negative response of a particular 2-D Gabor filter having a certain size, symmetry, and position on the iris. This process is shown in equations 4, 5, 6 and 7. Since it is the "sign bit" to be encoded, this information corresponds to the most significant bit (MSB) of the coefficient resulting from the integration of the product of the input image and the 2-D Gabor filter as described above. Extracting the independent information using the two quadrature symmetry and the odd quadrature symmetry of the 2-D Gabor filter is the corresponding real and imaginary parts of the complex 2-D Gabor filter. The bits determined by are denoted by the subscripts Re and Im.

These conditional expressions (Equations 4 through 7) can be used to determine the multiple scale (parameter) , , All sampled positions (polar parameter r and Each 2,048 bits in the iris code. Notably, almost all of the data compression is achieved in such code because of its decorrelating nature. The original iris image may typically consist of 262,000 bytes (512 × 512 pixel arrays, each pixel requires one byte), while most iris textures are only 1 by this multi-scale 2-D Gabor code. It was reduced to a very compact sine consisting of only about 1,000 data (ie 256 bytes).

An example of a 256-byte iris code is written at the top of FIG. 2, consisting of 256 columns, each with 8 bits calculated in the concentric analysis band. Some code has 2,048 bits, but the code has independent binary degrees of freedom less than 2,048. The main reason is the presence of significant radial correlation in the iris. For example, some furrows tend to spread over significant radial distances, affecting some distant parts of the cord. The second reason is that the correlation is introduced by the lowpass components of the bandpass 2-D Gabor filter. In particular, any signal convolved with a linear filter will obtain a correlation distance equal to the reciprocal of the filter's bandwidth.

Indeed, a large number of independent degrees of freedom can be estimated by examining the distribution of hamming distances (parts of inconsistent bits) over very many iris codes and comparing each code with every other code computed from the other iris for each bit. Since each bit has the same probability of 1 or 0, there is a probability p = 0.5 that a pair of bits from different iris codes will not match. When each 2,048 bits in a code are independent of every other bit, the distribution of observed hamming distances is equivalent to a binomial distribution with p = 0.5 and N = 2,048 (in other words, repeating the intact coin Equivalent to counting the top part of each throw of a 2048 coin toss). The actual distribution of the Hamming distance observed in the 2064 codes note from the other iris is shown in FIG. Its standard error is sigma = 0.038 in the vicinity of the average μ = 0.502. The standard deviation of the binomial distribution Since the distribution of observed hamming distances is equivalent to a binomial distribution with N = 173 bits, it represents such a good match that is actually suitable for the observed iris code data shown in FIG. As such, the 2048 bit iris code has about 173 independent binary degrees of freedom.

Using a binary estimate of N = 173 binary degrees of freedom as a measure of the complexity or dimension of a 2048-bit iris code, one can calculate the probability that two codes from different irises will coincide. Since the 2-D Gabor filter has no positive or negative bias, the a priori odds that any given bit will be 1 or 0 are even, so the probability is two corresponding bits in two different iris codes. Is equal to 0.5. When factoring with partial correlation within the iris code and assuming independence between the iris codes, there are 2 ¹⁷³ even numbers where two different irises will produce the same code, equal to one of 10 ⁵² .

The process of comparing any two iris codes, that is, previously stored code (block 22) and currently calculated code (block 24) from the donor phase (Figure 1, block 26), is a general format and constant length of all such codes. Because it is very simple. A similarity metric, called a hamming distance, is calculated to measure the "distance" or similarity between two codes. This measurement simply adds the total number of times when two corresponding bits in the two iris codes are inconsistent. Since the 2048 corresponding bit pairs all match, the Hamming distance between any iris code represented as a fragment between 0 and 1 and its own exact copy will therefore be zero. The Hamming distance between any iris code and its number (餘 數: all bits are just inverted) is 1. The Hamming distance between two random and independent bit streams is expected to be 0.5, since the corresponding random bit pairs have a 50% probability of matching and a 50% probability of inconsistency. In other words, if two iris codes originate from different eyes, their hamming distance will be 0.5 and in other cases their hamming distance is expected to be considerably lower. If two iris codes are calculated from the same picture, their hamming distance should be close to zero.

Comparisons between the iris cords can be made with several different relative shifts along their angular axis to compensate for possible subject's head or torsional eye rotation. These relative transitions in the code comparison process are implemented directly by lateral scrolling of the iris code relative to each other, although the code shown in the upper left corner of FIG. 2 is surrounded by a cylinder, combining the left and right margins and then the cylinder. The rotation is repeated to repeat the comparison process.

The calculation of the Hamming distance between iris codes is done very simply by using the basic logical operator XOR (exclusive-OR). A pair of bits A and B can have exactly four possible combinations: (AB) = (00), (01), (10) and (11). The XOR operator for two inputs is defined as 1 if only one of the inputs is one; Otherwise their XOR is zero. As such, in the given example for four possible combinations of bits A and B, the corresponding values of their XOR are (A XOR B) = 0,1,1,0. Clearly, XOR can thus be used to detect inconsistencies between any pair of bits, whatever their value.

Computing the total number of times when the XOR of two corresponding iris code bits equals 1 and dividing by the total number of such comparisons (number of bits in the iris code) is equivalent to measuring the Hamming distance between the two codes. Instead, this amount may also be described as the normalized squared length or squared norm of the difference vector between two iris code vectors in binary space in the 2048 dimension. Both of these formulas produce the same measure of iris code comparison, and they contribute directly to converting themselves with a calculated probability that two iris codes come from the same person because they come from the same iris.

The problem of recognizing a given sign of an iris as belonging to a particular individual or determining that he / she is a fake can be formalized within the foundation of statistical pattern recognition and decision theory.

Yes / no decisions in pattern recognition have four possible outcomes: whether or not a given pattern is the real one of the categories that are characterized; And whether the decision made on either of these two cases is correct or incorrect, these four results are commonly referred to as Hit, Miss, False Alarm and Correct Rejection. Four possible outcomes in this application are Real Acceptance of Authentic (AA), Acceptance of Imposter (IA), Real Rejection of Authentic (AR) and Rejection of Imposter (IR). The goal of the decision-making algorithm is to maximize the probabilities of AA and IR while minimizing the probabilities of IA and AR. Paired trade-offs among the probabilities of these four results can be done in a way that reflects the costs and benefits associated with them in a particular application.

The formulation of the decision-undr-uncertainty is provided in Figure 6. A given measurement of a Hamming distance, or a fragment of inconsistent bits between two iris codes, constitutes a point in abscissa. The measurement is treated as a random variable describing one of two processes represented by two overlapping probability distributions. Which of these two distributions is deductively unknown to what this random variable describes; This goal is to determine this. The criterion is selected in FIG. 6 by the vertical dotted line so that all hamming distances smaller than this criterion are considered to belong to the "real" distribution while all hamming distances larger than this criterion are considered to belong to the "fake" distribution. These two distributions, P _AU (H) and P _IMP (H), represent the specially measured Hamming distances resulting from each of the two comparisons of the same iris ("real") or from each of the two comparisons of the other iris ("fake") Gives a probability density of H).

Four results, AA, IA, AR and IR now have a probability completely determined by the selected criterion and also by the statistical parameters of the distribution that are the two criteria. Decision rules

Accept if Hamming distance <criterion

Hamming distance> reject if standard

The probability of four possible outcomes is equal to the area under two probability density functions, P _AU (H) and P _IMP (H) on both sides of the selected host (c).

These four probabilities are represented by four regions lined in FIG.

The four probabilities are divided into 24 pairs that add up to 1, and these two pairs are governed by inequality:

P (AA) + P (AR) = 1 (12)

P (IA) + P (IR) = 1 (13)

P (AA)> P (IA) (14)

P (IR)> P (AR) (15)

In addition, the two error rates P (AR) and P (IA) are minimized if the two Hamming distance distributions P _AU (H) and P _IMP (H) have minimal overlap. This can be accomplished by putting their two means further away or by reducing their dispersion, or both. In general, it is noted that these two distributions do not have the same shape and dispersion for simplicity as implied in FIG.

The usefulness or discernment of the biometric signature method for discrimination and recognition between individuals can be defined by the amount of overlap between these two distributions. Clearly, in the absence of overlap, 100% accurate decisions can be made. Conversely, the more overlap, the greater the rate of error regardless of the decision criteria used.

By defining decision criteria C in equations (8)-(11) above, the ability to select different decision strategies that are most suitable for different applications is provided. For example, in controlling access to a bank account from an automated teller machine, allowing ARR to be much higher than zero can be bad for customer relationships, even if it means tolerating higher IARs. ; After all, the cost of accepting a fake is in the worst case the ATM withdrawal limit. On the other hand, in the case of military or embassy security systems, fakes are admitted, although in this case it would be necessary to tolerate higher ARR (the true percentage rejected for further screening). Much more conservative standards were required to be as strict as once a million times.

The processing of decision criterion C to implement another decision strategy is briefly described in FIG. Theoretical Authentic Acceptance Rate, or P (AA), is the theoretical Imposter Acceptance Rate as a trajectory of points determined by different choices for decision criteria C shown in FIG. It is plotted against P (IA). These two forms are only theoretical to clarify the nature of the crystal problem; As shown, they suggest greater uncertainty than what is actually present for this biometric recognition system.

Equation (14) tells us that the strategy curve shown in Figure 7 is always above the diagonal in this probability space. In general, we try to use a decision strategy that generates points as close as possible to the upper left corner, because reaching that typical means that all reals are accepted while all fakes are rejected. Obviously, overly conservative or overly free strategies correspond to sliding down a curve toward two diagonals, in which case the real and the fake are either totally rejected (bottom) or totally accepted (right). Obviously, in FIG. 7 the identification power will be zero anywhere along the diagonal and will be equal to 1 in the upper left corner of the space. Therefore, regardless of the location chosen to place the decision criteria along the free-repair strategy curve, the overall capability of the detection method is the length of the line that meets the bend in the diagonal and the strategy curve, ie the "arrow" in the "bow." It can also be measured by the length of the arrow.

By formalizing biometric identification problems within the framework of signal processing and statistical decision theory, we can now value people's cognitive abilities by their irises.

The distribution of Hamming distances calculated among the 2064 pairs of different irises (labeled "fake" because no pairs came from the same person) is already shown in FIG. As expected, the average hamming distance is close to 0.5 because the probability of matching or inconsistent any bit in a 2048-bit (256 byte) code for two different irises is the same. The distribution of hamming distances is concentrated around the expected value; The actual mean is μ = 0.497 and the standard deviation is σ = 0.038.

8 shows the distribution of calculated Hamming distances among 1208 pairs of different images of the same iris ("real") obtained at different times. Ideally, these Hamming distances should be zero if the phases are really the same; However, differences in the angle of gaze, partial eyelid blockage, specular reflection from the retina and relative contraction of the pupils result in some differences in the encoded structure. Nevertheless, these hamming distances are clearly substantially smaller than the fake shown in FIG. The true distribution has a mean of μ = 0.084 and a standard deviation of σ = 0.0435.

Due to the cycloversion of the eye in the socket and the torsional rotation that might otherwise occur when the person tilts the head, it is necessary to compare all iris codes over a range of different state orientations. . The best agreement obtained from a combination of such comparisons is only retained as a measure of similarity. Since such a "best of n relative orientation" test always selects the lowest Hamming distance for both real and fake comparisons, these distributions all transition to the left and are narrower than otherwise. Of course, this does not affect the nature of the decision work and is not based on any assumptions about the form of the two distributions. However, it improves the overall decision performance by gaining more benefit from comparing those in the relative orientation of several candidates than in matching between codes for the iris, regardless of the degree of matching between different images of the same eye.

The true distribution shown in FIG. 8 was obtained according to the best rule of seven orientations. Figure 9 shows this distribution in relation to its corresponding counterpart to the fake, which compares the same set of 2064 irrelevant iris codes that were used for Figure 5 but now has the same new "7 orientations." Best of all "rule. These two distributions do not have any experimental overlap. FIG. 10 shows the same histogram pair adjusted by theoretical binomial distribution, whose parameters are adjusted to match the mean and variance of the observed experimental distribution pairs. The binomial form is mathematically appropriate given the nature of the code comparison process as a statistical sequence in the Bernoulli trial. The problem of iris pattern recognition thus translates into a statistical test essentially of independence. Exclusive OR tests between iris code bits test the hypothesis that two code sequences can result from independent random processes. This test of statistical independence almost certainly fails for two codes from the same eye, but almost certainly for two codes from different eyes.

Although the two experimental distributions shown in FIG. 9 have no overlap and actually fall of the observed points in the range of 0.25 to 0.35 hamming distance, there is a slight gap between the two distributions when studying a theoretically large enough database. There is an overlap of. The adjusted pair of theoretical binomial curves superimposed in FIG. 10 provides a method for estimating the error rate when there are an unlimited number of observations. These overlapping adjusted distributions should be taken into account when considering FIG. 6, which provides a basic basis for statistical decisions. As defined in equations (8)-(11), the probability of personal identification or de-identification and the predicted error rate can be calculated as cumulative integral under these two distributions on either side of any selected Hamming distance determination criteria. Can be.

In the comparison of the observed population of 1208 pairs, the error rate to false accept (IAR) and the true rate to reject (ARR) are plotted as a function of the Hamming distance criterion in FIG. This graph clarifies the exchange between Type I and Type II error rates (IAR and AAR) that can be addressed by changes in the standard. Their theoretical intersection occurs for a Hamming distance criterion of about 0.321, where the probability of false acceptance and false rejection is both 131,000th. For example, if a more conservative decision criterion is required, such as a Hamming distance of 0.26, that is, the probability of accepting fakes is 2 billion minutes a day, Figure 11 correctly accepts the true rate of 99.96%. It is showing much higher. The dotted line is the theoretical error rate calculated according to equations (7) and (8) using the adjusted binomial distribution for P _AU (H) and P _IMP (H) shown in FIG.

The previously developed binomial basis for this statistical pattern recognition work allows us to calculate the level of confidence associated with any decision to verify or deny the person's identity based on their iris signature. In particular, when measuring their Hamming distances by comparing two iris codes, eg, those previously "enrolled" with the present, a Hamming distance of this size or smaller occurs by chance from two different irises. We can calculate the probability Only when this probability is small enough will the person be accepted as genuine.

The confidence level associated with the decision is that p is the probability that a pair of bits do not match, and if m is the probability that they match, then m ≤ CN bit mismatches (thus the Hamming distance HD is less than or equal to fragment C of the mismatch). All possible combinations of N independent binary degrees of freedom with You need to sum up the product of the probabilities of each such event at:

Equation (16) defines the probability that the Hamming distance HD between codes for different irises may be accidentally smaller than a given decision criterion C. This allows us to determine how likely there is a probability of responding to an opportunity for false acceptance for any given decision criterion. [Stirling's approximation allows us to estimate the large factorial estimates needed to evaluate equation (16).] As shown in the table in Figure 12, these theoretical probabilities are 0.30. It is generally about 2.4 million for Hamming distance criteria and this probability quickly reaches a "planetary" level (of 1 billion) for 0.26 or lower criteria. Obviously, using such criteria, such unlikely errors will occur in existing databases. Indeed, the false acceptance rate observed in 2064 iris code comparisons is already at zero for a Hamming distance criterion as high as 0.35, at which point the theoretical true rejection rate is about one millionth (end line in Figure 12).

The final dimension of the capability of the biometric sign security system of the present invention is the calculated confidence level associated with the normal or average hamming distance encountered in different phases of the same iris. As evident in the bar graphs of FIGS. 8 and 9, the average hamming distance between two iris codes originating from the same iris was 0.084. In this general case, the level of confidence in the decision to really accept the individual is actually astronomical. In particular, the probability that a hamming distance of this average size or less will result from a fake is one tenth ^thirty minutes according to equation (16).

The analysis performed by this embodiment of the present invention generates an identification code of 256 bytes (2048 bits) of data. This number has been found to generate reliable identification codes in combination with optimized processing characteristics. This iris code maps all other irises to a general abstract mathematical code of constant length. This allows code comparisons to be as efficient as possible even between relatively "featureless" irises (probably as a result of low contrast shooting) and those with rich visual textures. .

This comparison process itself contributes directly to a simple hardware mechanism based on the XOR gate. An XOR gate consisting of just three transistors is available on standard semiconductor chips. For example, a standard IC called 74F86 includes four independent XOR gates that can operate at 80 MHz, and are available at very low cost. The comparison between two iris codes and the overall retrieval of a large database of stored iris codes can be implemented very quickly and essentially in parallel. For example, a circuit board containing a 32 X 32 array of 74F86 ICs would compare the current iris code against a population of 160 million previously stored iris codes in less than one second, allowing the individual to use one of them. It is possible to establish whether the recognition is reliable.

Since each bit in the code can be thought of as a binary random variable, the theory of binomial statistics applies to estimating the probability that any given fragment of a bit in two different iris codes will coincide. This allows you to use statistical decision theory to make an overview of confidence levels for all decisions. So far, as long as the pattern recognition problem is transformed here into a statistical test of independence based on the polarity of Gabor coefficients computed across multiple analysis scales in a dimensionless coordinate system, the overall statistical basis of the present invention is the traditional decision theory and unique signal processing method. It can be considered as a synthesis of.

It should be understood that the above embodiments are merely exemplary of the present application. Other embodiments will be readily devised by those skilled in the art within the spirit and scope of the present invention.

Claims (23)

In the method of differentiating and identifying a specific person by biometric analysis of the iris of the eye, Obtaining an image of the eye of the person to be identified; Separating and defining the iris of the eye within the image, the step comprising: defining a circular pupil boundary between the iris portion and the pupil portion of the image; Defining another circular boundary between the iris portion and the sclera portion of the phase using an arc that is not necessarily centered with the pupil boundary; The isolated iris is a polar coordinate system whose origin is at the center of the circular pupil boundary and whose radial coordinate is measured as a percentage of the distance between the circular pupil boundary and the circular boundary between the iris and the sclera. Setting for the phase; And defining a plurality of annular bands within the iris phase; Analyzing the iris to generate a presenting iris code; Comparing the indication code with a previously generated reference iris code to generate a measure of the similarity between the indication iris code and the reference code; And And converting the similarity measure to determine whether the iris codes originate from the same iris. 2. The method of claim 1, further comprising calculating a confidence level for the determination. The biometric person based on iris analysis according to claim 1, wherein the analysis zone excludes any preselected portions on the iris that may be occluded by mirror reflections from the eyelids, eyelashes or illuminator. Identification method. 4. The method of claim 3, wherein said analyzing step comprises: analyzing an iris image portion within said annular analysis bands; Generating an iris code for said iris phase portion using signal processing means. 5. The method of claim 4, wherein said signal processing means comprises applying multi-scale, self-similar, two-dimensional quarter pass filters from polar coordinates onto the iris. Biometric person identification method based on iris analysis. 6. The method of claim 5, wherein the iris code has a certain number of bits and a universal format for all irises. 8. The method of claim 7, wherein the analyzing step removes luminance bias, removes slow luminance gradient resulting from oblique illumination, removes noise, and prevents aliasing. Applying a band pass filter to the region of the original iris phase signal. 8. The method of claim 7, wherein the value of each bit in the iris code is defined as: A biometric based iris analysis characterized in that it is defined as "1" or "0" by calculating the most significant bit of the filter output for any region of the iris that makes support for the filter of a given size at a given location. Measurement person identification method. The method of claim 8, wherein the comparing step, Calculating a basic logical XOR (exclusive OR logical operation) between all corresponding bits to compare any two iris codes; And Calculating a squared norm of the resulting binary vector, And the comparison value is defined as a hamming distance between the two iris code vectors. The method of claim 9, wherein the subject's head tilts or cyclovergence; That is, the bit further comprises the step of repeating the comparison step for several different relative transitions of the iris code along each angular axis to correct for the possibility of torsional eye rotation. Biological specific person identification method based on iris analysis. 11. A method according to claim 10, wherein the calculating step comprises the step of transforming the Hamming distance to calculate the likelihood that two codes originate from the same iris, i.e. the same person. . 12. The method according to claim 11, wherein the calculated likelihood that the observed matching fraction of bits in the codes may coincide if the indication code and the reference code are independent, i.e., even if they originate from separate irises. A biometric person identification method based on iris analysis, characterized in that it is found by calculating the probability of the presence. 13. The method of claim 12, wherein the measured hamming distance is converted to a probability that the two iris codes originate from the same eye; A preselected criterion is applied to the measured hamming distance to generate a “yes” or “no” decision; The confidence level for the decision is given by the calculated probability Biometric person identification method based on iris analysis characterized by the above-mentioned. The method of claim 14, The circular pupil boundary is defined by the following definition: Where r is the radius of the boundary, x ₀ and y ₀ are the center coordinates, I is the image intensity, And said radius and center coordinates change in a predetermined pattern. The method of claim 14, wherein the other circular boundary between the iris portion and the sclera portion of the phase is: Is defined by determining the distance from the coordinate system origin to the left and right limbs, where r is the previously defined pupil radius, Is the radial shell distance, Is the original image intensity in polar coordinates, To find the limbus at the meridian at 3 o'clock or 9 o'clock, respectively. Biometric person identification method based on iris analysis, characterized in that the equivalent to. 16. The method of claim 15, wherein the analysis bands comprise a plurality of annular bands fully spread around the pupil and a plurality of semi-cyclics having polar coordinate angles around the pupil from about 45 degrees to 135 degrees and from 225 degrees to 315 degrees. Biometric person identification method based on iris analysis characterized by including the band. 17. The method of claim 16, wherein the analysis bands are spaced at equal radius distances from an internal point some distance preselected from the pupil boundary to an external point some distance preselected from the rim booth, the interval being the coordinates. Biometric person identification method based on iris analysis, characterized in that the weighted at an angle considering the difference between the fractional distance from the center to the left and right limbs. 15. The method of claim 13, wherein the band pass filter is a 2-D Gabor filter. 19. The method of claim 18, wherein the 2-D Gabor filter is defined in polar coordinates as follows: Where r is the radius, Is the distance in radians, Is the frequency, Wow Biometric person identification method based on iris analysis, characterized in that the constant. 20. The method of claim 19, wherein the iris code consists of 2048 bits. The circular pupil boundary of claim 3 is defined by the following relationship: Where r is the radius of the boundary, x ₀ and y ₀ are the center coordinates, I is the phase intensity, And wherein the radius and center coordinates are systematically changed by iterative gradient ascent to find the maximum value in the defined relationship. The method of claim 21 wherein the other circular boundary between the iris portion and the sclera portion of the image is: Is defined by determining the distance from the coordinate system origin to the left and right limbs, where r is the previously defined pupil radius, Is the radial shell distance, Is the original image intensity in polar coordinates, To find the limbus at the meridian at 3 o'clock or 9 o'clock, respectively. Biometric person identification method based on iris analysis, characterized in that the equivalent to. An apparatus for differentiating and identifying a specific person by biometric analysis of an iris of an eye, An imaging device for obtaining an image of the eye of the person to be identified; And Include processors, The processor separates and defines the iris of the eye within the image, defines a plurality of annular analysis bands within the iris, analyzes the iris to generate a presenting iris code, and displays the display code. Generate a measure of the similarity between the indication iris code and the reference code compared to a previously generated reference iris code, determining that the iris codes are from the same iris or not A biometric person identification device based on iris analysis, characterized in that it is programmed to transform a similarity measure.