Iris Modeling and Synthesis

1. Iris Modeling and Synthesis CPSC 601 Biometric Technologies Course

2. Lecture Plan • Motivation • Iris structure • Iris Image acquisition • Methodology • Iris Localization • Iris features • Matching • Iris Synthesis • Future Developments

3. Iris Structure Anatomy of the human iris. The upper panel illustrates the structure of the iris seen in a transverse section. The lower panel illustrates the structure of the iris seen in a frontal sector.

4. Iris Structure • At a finer grain of analysis, the iris is composed of several layers. The posterior surface is composed of heavily pigmented epithelial cells that make it impenetrable to light. Anterior to this layer two muscles are located that work in cooperation to control the size of the pupil. The visual appearance of the iris is a direct result of its multilayered structure. Iris color results from the differential absorption of light impinging on the pigmented cells in the anterior border layer.

5. Iris Structure - Uniqueness The first source of evidence comes from clinical observations. During the course of examining large number of eyes, ophthalmologists have noted that the detailed spatial pattern of an Iris seems to be unique. The pattern seem to vary little, at least past childhood. The second source of evidence comes from developmental biology. While the general structure of the iris is genetically determined, the particulars of its minutiae are critically dependent on circumstances (e.g. the initial conditions in the embryonic precursor to the iris). Anatomy of the iris visible in an optical image.

6. Iris Structure - dynamics Another interesting aspect of the physical characteristics of the iris from a biometric point of view has to do with its dynamics. Due to the complex interplay of the iris's muscles, the diameter of the pupil is in a constant state of small oscillation at a rate of approximately 0.5 Hz. This movement could be monitored to ensure that a live specimen is being evaluated. Since the iris reacts very quickly to changes in impinging illumination, monitoring the reaction to a controlled illuminant could provide similar evidence.

7. Iris Image Acquisition Acquisition of a high-quality iris image, while remaining non-invasive to human subjects, is one of the major challenges of automated iris recognition. Figure: Passive sensing approaches to iris image acquisition. The upper diagram shows a schematic diagram of the Daugman image acquisition rig. The lower diagram shows a schematic diagram of the Wildes et al. image acquisition.

8. Iris Image Acquisition In order to cope with the inherent variability of ambient illumination, extant approaches to iris image sensing provide a controlled source of illumination as a part of their method. Research initiated at Sarnoff Corporation and subsequently transferred to Sensar Incorporated for refinement and Commercialization has yielded the most non-invasive approach to iris image capture that has been documented to date. For capture, a subject merely needs to stand still and face forward with their head in an acquisition volume of 600 vertical by 450 horizontal and a distance of approximately 0.38 to 0.76 m, all measured from the front-center of the acquisition rig. Capture of an image that has proven suitable to drive iris recognition algorithm can then be achieved totally automatically, typically within 2-10 seconds.

9. Iris Image Acquisition Figure: Active sensing approach to iris image acquisition.

10. Iris Image Localization • Following image acquisition, the portion of the image that corresponds to the iris needs to be localized from its surroundings. The iris image data can then be brought under a representation to yield an iris signature for matching against similarly acquired, localized and represented irises.

11. Iris Image Localization Daugman and Wildes et al. approaches make use of first derivatives of image intensity to signal the location of edges that correspond to the borders of the iris. Here, the notion is that the magnitude of the derivative across an imaged border will show a local maximum due to the local change of image intensity. Both systems model the various boundaries that delimit the iris with simple geometric models. For example, they both model the limbus and pupil with circular contours. The Wildes et al. system also explicitly models the upper and lower eyelids with parabolic arcs. In initial implementation, the Daugman system simply excluded the upper and lower most portions of the image where eyelid occlusion was most likely to occur; subsequent refinements include explicit eyelid localization.

12. Iris Image Localization • The two approaches differ mostly in the way that they search their parameter spaces to fit the contour models to the image information. The Daugman approach fits the circular contours via gradient ascent on the parameters (xc, yc, r) so as to maximize where is a radial Gaussian with center r0 and standard deviation σ that smoothes the image to select the spatial scale of edges under consideration, * symbolizes the convolution, ds is an element of circular arc and division by 2Πr serves to normalize the integral.

13. Iris Image Localization • The Wildes et al. approach performs its contour fitting in two steps. First, the image intensity information is converted into a binary edge-map. Second, the edge points vote to instantiate particular contour parameter values.

14. Iris Image Localization Both approaches to localizing the iris have proven to be successful in the targeted application. The histogram-based approach to model fitting should avoid problems with local minima that the active contour model's gradient descent procedure might experience. However, by operating more directly with the image derivatives, the active contour approach avoids the inevitable thresholding involved in generating a binary Edge map. Illustrative results of iris localization. Given an acquired image, it is necessary to separate the iris from the surroundings. Taking as input an iris image, automated processing delineates that portion which corresponds to the iris.

15. Iris modeling- methodology • The distinctive spatial characteristics of the human iris are displayed at a variety of scales. • The Daugman approach makes use of a decomposition derived from application of a two-dimensional version of Gabor filters to the image data. Since the Daugman system converts to polar coordinates, (r, θ), during matching, it is convenient to give the filters in a corresponding form as where and co-vary in inverse proportion to generate a set of quadrature pair frequency selective filters, with center locations specified by (r0, θ0). These filters are particularly notable for their ability to achieve good joint localization in the spatial and frequency domains.

16. Iris modeling- methodology • The Wildes et al. approach makes use of an isotropic bandpass decomposition derived from application of Laplacian of Gaussian (LoG) filters to the image data. The LoG filters can be specified via the form with σ the standard deviation of the Gaussian and ρ the radial distance of a point from the filter’s center. In practice, the filtered image is realized as a Laplacian pyramid.

17. Iris modeling- methodology By retaining only the sign of the Gabor filter output, the Representational approach that is used by Daugman yields a remarkably parsimonious representation of an iris. Indeed, a representation with a size of 256 bytes can be accommodated on the magnetic stripe affixed to the back of standard credit/debit cards. In contrast, the Wildes et al. representation is derived directly from the filtered image for size on the order of the number of bytes in the iris region of the originally captured image.

18. Iris matching Iris matching can be understood as a three-stage process. • The first stage is concerned with establishing a spatial correspondence between two iris signatures that are to be compared. • Given correspondence, the second stage is concerned with quantifying the goodness of match between two iris signatures. • The third stage is concerned with making a decision about whether or not two signatures derive from the same physical iris, based on the goodness of match.

19. Iris matching Given the combination of required subject participation and the capabilities of sensor platforms currently in use, the key geometric degrees of freedom that must be compensated for in the underlying iris data are shift, scaling and rotation. Shift accounts for offsets of the eye in the plane parallel to the camera's sensor array. Scale accounts for offsets along the camera's optical axis. Rotation accounts for deviation in angular position about the optical axis. Another degree of freedom of potential interest is that of pupil dilation.

20. Iris matching • Daugman’s system uses radial scaling to compensate for overall size as well as a simple model of pupil variation based on linear stretching. The scaling serves to map Cartesian image coordinates (x, y) to polar image coordinates (r, θ) according to where r lies on [0, 1] and θ is cyclic over [0, 2Π], while (xp(θ), yp(θ)) and (x1(θ), y1(θ)) are the coordinates of the pupillary and limbic boundaries in the direction θ. Rotation is compensated for by brute force search: explicitly shifting an iris signature in θ by various amounts during matching.

21. Iris matching • The Wildes et al. approach uses an image registration technique to compensate for both scaling and rotation. This approach geometrically projects an image, Ia(x, y), into alignment with a comparison image, Ic(x, y), according to a mapping function (u(x, y), v(x, y)) such that, for all (x, y), the image intensity value at (x, y) – (u(x, y), v(x, y)) in Ia is close to that at (x, y) in Ic. More precisely, the mapping function (u, v) is taken to minimize while being constrained to capture a similarity transformation of image coordinates (x, y) to (x’, y’), i.e. with s a scaling factor and R(Φ) a matrix representing rotation by Φ.

22. Iris matching An appropriate match metric can be based on direct point wise Comparisons between primitives in the corresponding signature representations. The Daugman approach quantifies this matter by computing the percentage of mismatched bits between a pair of iris representations, i.e. the normalized Hamming distance. Letting A and B be two iris signatures to be compared, this quantity can be calculated as With subscript j indexing bit position and denoting the exclusive-OR operator. The Wildes et al. system employs a somewhat more elaborate procedure to quantify the goodness of match. The approach is based on normalized correlation between two signatures (i.e. pyramid representations) of interest.

23. Iris matching The final subtask of matching is to evaluate the goodness of match values to make a final judgement as to whether two signatures under consideration do (authentic) or do not (impostor) derive from the same physical iris. In the Daugman approach, this amounts to choosing a separation point in the space of (normalized) Hamming distances between the iris signatures. Distances smaller than the separation point will be taken as indicative of authentics; those larger will be taken as indicative of impostors. In the Wildes et al. approach, the decision making process must combine the four goodness of match measurements that are calculated by the previous stage of processing (i.e. one for each pass band in the Laplacian pyramid representation that comprises a signature) into a single accept/reject judgement.

24. Iris modeling – future developments Further developments could be focused on yielding more compact Systems that can be easily incorporated into consumer products where access control is desired (e.g. automobiles, personal computers, various handheld devices). • Can iris recognition be performed at greater subject to sensor distances while remaining unobtrusive? • How much subject motion can be tolerated during image capture? • Can performance be made more robust to uncontrolled ambient illumination? Toward iris recognition at a distance. An interesting direction for future research in iris recognition is to relax constraints observed by extant systems. As a step in this direction, an iris image captured at 10m subject to sensor distance is shown.

25. Iris modeling – future developments At a more operational level of performance analysis, studies of iris recognition systems need to be performed wherein details of acquisition are systematically manipulated, documented and reported. Parameters of interest include, geometric and photometric aspects of the experimental stage, length of time monitored and temporal lag between template construction and recognition attempt. Similarly, details of captured irises and relevant personal accessories need to be properly documented in these same studies (e.g. eye color, eyewear).

26. Iris Synthesis • Classical Biometrics - Recognition • Fingerprints, Faces, Irises • Inverse Problem Synthesis • Testing Recognition methods

27. Iris Synthesis - Goals • Synthesis Of Biometric Databases • Iris Database Augmentation • Testing Recognition Methods • Minimal User Input

28. Iris Synthesis - previous work • Iris Recognition - [Wildes 94, Daugman 04] • Biometric Synthesis - [Yanushkevich et al. 04] • Iris Synthesis - [Lefohn et al. 03, Cui et al. 04]

29. Iris Synthesis • An Ocularists Approach to Human Iris Synthesis. • [ Lefohn et. al. 03] • An Iris image synthesis method based on PCA and Super-Resolution. • [Cui et. al. 04]

30. PCA Approach • Uses 75 Dimensional PCA Feature Vector • Randomization • Super Resolution • Great statistical results • Low Realism

31. Ocularists Approach • Uses: 30-70 Layers • Great Results. • Domain Specific Knowledge An ocularist's approach to human iris synthesis. Lefohn et. al. 2003. Used with permission.

32. New Approach • Use Real Iris Sample • Use sets of Similar Irises • Capture Characteristics • Chaikin Reverse Subdivision • Combine Characteristics • Multiple Iris Donors See L. Wecker, F. Samavati and M. Gavrilova works on the subject.

33. Reverse Subdivision Global & Local Features Linear Implementation Realistic results Comparison • PCA • Global Features • Not as Efficient • Realism

34. Organization

35. Method • First step: Isolate the iris. • Polar Transform • Iris Stretching

36. Multiresolution • Data has many resolutions • Levels of resolution have different meanings • Reverse Subdivision • Details

37. Decomposition

38. Method Courtesy of: Michal Dobes and Libor Machala, Iris Database, http://www.inf.upol.cz/iris/ • Capture Details • Reverse Subdivision • Details • All Characteristics

39. Combinations

40. Classifications • Frequency of Data • Number of Concentric Rings Courtesy of: Michal Dobes and Libor Machala, Iris Database, http://www.inf.upol.cz/iris/

41. Database Size

42. Original Set Input Irises Courtesy of: Michal Dobes and Libor Machala, Iris Database, http://www.inf.upol.cz/iris/

43. Output Irises Courtesy of: Michal Dobes and Libor Machala, Iris Database, http://www.inf.upol.cz/iris/

44. Combinations

45. Output Irises Courtesy of: Michal Dobes and Libor Machala, Iris Database, http://www.inf.upol.cz/iris/

46. Future Work • Post-Processing • Multiple samples of each iris • Verification • Statistically