Iris Recognition: Algorithms and Systems

Iris Recognition Algorithms and Systems

• Chapter focuses on iris recognition with a special focus on algorithms and recognition systems.

Iris Biological Structure

• The iris is an internal eye organ situated behind the cornea and in front of the lens.
• The iris is a multilayered structure.
• The posterior layer is at the back, is heavily pigmented, two cells thick, and blocks light.
• The muscle layer features sphincter and dilator muscles that contract and dilate the pupil.
• Located above the muscles, the stromal layer is composed of collagenous connective tissue and blood vessels arranged radially.
• The anterior border layer is the foremost layer, having a higher density of chromatophores (pigment cells) than the stromal layer.
• Anterior portion refers to the muscles, stroma, and border layers, that form visible part of the iris.
• The anterior portion is focused on automated iris recognition.
• The iris appears as an annulus, bounded by the pupillary boundary and the limbus boundary, when viewed by a near-infrared camera.
• Central Zone: pupillary area.
• Peripheral Zone: ciliary muscles
• The two zones are split by the collarette, a circular zigzag ridgeline.
• Crypts (Fuchs crypts) are small pit-like, irregular features in the region of the collarette, that allow fluids to enter and exit as the pupil contracts.
• The iris of the eye has a complex structure with a rich texture.
• The term "texture" characterizes an image's homogeneity, coarseness, and directionality
• Iris texture differs substantially across the population.
• The morphogenesis of the texture relief, which occurs randomly during prenatal development, is why each iris is unique.
• Even monozygotic twins exhibit differences in patterns, implying epigenetic, random control.
• Melanin determines the Iris's color.
• The color of the iris is not important for identification of iris.
• The texture detail present in the anterior portion of the iris is useful for recognition.

Iris Recognition Applications

• Iris recognition is used at airports for passengers, employees and flight crews, requiring rapid processing and high accuracy, particularly against watch lists.
• The iris recognition system in the UAE is used to identify expellees trying to re-enter.
• Iris recognition can be used in coal mines, especially when individuals can't give good face or fingerprint data because of the surroundings.

Modules of an Iris Recognition System

• The iris recognition system matches patterns by comparing irides and giving a score reflecting similarity/dissimilarity.
• A typical system has four parts:
  • Image Acquisition
  • Iris Segmentation
  • Iris Normalization
  • Iris Encoding and Matching
• The four modules of an Iris Recognition System are shown
• Matching performance relies on accurate segmentation, to correctly establish the iris's spatial extent.

Iris Image Acquisition

• Getting a 2D eye image is the goal, using a monochromatic CCD camera, sensitive to near-infrared spectra.
• An external NIR (near-infrared) source, often paired with the acquisition system, lights the iris.
• Cooperative individuals who place their eye close enough to the camera are required for most systems.
• Generally, a series of captures are taken, and images are maintained if they have sufficient iris texture quality.
• A sensor with the ability to pick up near-infrared (NIR) radiation is used to get the iris image.
• The specific range is generally about 700-900nm in the IR spectrum.

Iris Segmentation

• Isolation of the iris from the eye is the function of the segmentation module.
• Surrounding features include : sclera, pupil, eyelids, and lashes
• Segmentation typically involves
  • Inner/outer iris boundaries; namely pupillary and limbus boundaries
  • Identifying eyelashes and eyelids that interrupt limb us boundary.
• An integro-differential operator detects iris boundaries.
• Accurate iris segmentation is vital, or matching accuracy suffers.

Iris Normalization

• Geometric normalization maps the iris once inner and outer boundaries get estimated.
• The iris texture transforms from Cartesian to pseudo-polar coordinates using a "rubber sheet" model, to transform the iris texture.
• The unwrapping of the iris results in a rectangular shape.
• Rows reflect angular direction and columns represent radial direction.
• The normalization is threefold::
  • Corrects for changes in pupillary extent.
  • Normalizing guarantees irides of multiple people map into a single image.
  • A simple translation during the matching stage accounts for head tilting during image acquisition.
• Each binary mask labels valid iris pixels "1," separating them from eyelid/eyelash pixels, marked "0."

Encoding and Matching

• Feature extraction methods encode texture.
• Irides can be compared when unwrapped through correlation filters.
• A multi-resolution analysis results from the encoding mechanism.
• Algorithms generally apply wavelet filters to the response data.
• Quadrature 2D Gabor Wavelets are commonly used to extract the local phasor information of the iris texture.
• Using the quadrant, each phasor response encodes with two piece of binary information.
• The term iris code is the binary code product
• Techniques can compare iris codes

Design of an Iris Recognition System:

• Iris recognition system design.

Iris Image Acquisition Notes:

• The use of Near-Infrared (NIR Light) has at least two clear advantages:
  • Nuances in dark iris colors can be resolved
  • Melanin properties cause textural nuances of dark irides not to be clearly resolved by visible light
  • Enhanced image with lower wavelengths
  • Elicit better details with NIR and sensor combinations
• NIR light is non intrusive
  • Because it can't be seen by the human eye
  • Minimizes the amount of NIR required, non-harmful
• Multispectral cameras captures red, green, blue, and NIR spectra
• Darker textural details are more evident in the NIR than in the red, blue, or green channels. (Figure 5-4)
• Human assessment reveals better texture in the NIR channel.
• Even dark irides benefit from NIR.
• Many setups take NIR images and will utilize subset quality standards.
• Good processing requires 100 - 200 pixels radial

Iris segmentation

• Occlusion can result from:
  • The image being impacted by many factors, these include closed eyelids.
  • Intruding eyelashes, non-uniform and harsh illumination, low resolution, pupil dilation or constriction.
• Capturing the image includes the pupil, sclera, lids and lashes.
• Iris is localized with spatial extent of these structures in a 2D eye image through localization or segmentation.
• The primary task is setting pixels that correspond with the iris.
• Segmentation presents a few challenges:
  • Texture changes across eyes causing stochastic textures or "edges" that are dispersed.
  • Lack of image models for content thereby precluding models (based on eye texture).
  • Shape is annular: Pupil, Sclera and Eyelids are internal perimeter
  • Inaccurate measurements of contours are a problem
  • In images there needs to be clear boundary estimations
  • Eyelids can affect the shape of estimated boundaries
  • Lashes may partly hide texture
  • Can produce spurious edges and affect segmentation

Segmentation using integro-differential operator

• The integro-differential operator is qualified by the order statistic (max)
• I is the input eye image and I(x,y) is the pixel intensity
• The image has a radial Gaussian filter G (r) and is convolved by scale σ to it
• The gradient of those circumference pixels is measured applying a circle of radius r and centered at (x,y)
• For pixels in picture, there is a computed gradient
• The contour of the pupillary boundary defines (r,xₒ,yₒ) in equation 4.1
• A similar procedure is also used/required to detect the limbus boundary
• The arc of integration will therefore constrain near-vertical pixels.
• Eye Images, boundaries and contours can all be affected

Iris Normalization Design Points:

• Many factors can impact the spatial extent of iris texture
• These factors can include:
  • Pupil dilation and contraction
  • Resolution of image/sensor and imaging distance
• Segmented Iris can be unwrapped from cartesian to normal polar coordinates.
• By implementing the Daugman's Rubber Sheet model, polar coordinates in the angular area can all transform correctly (see figure 5.6)
• Image remapping to cartesian remapping from polar coordinates via the remapping formula (5-2)
• Noise Masks can also facilitate

Iris encoding and matching

• This is the set of numerical features extracted from the iris.
• The process corresponds to the feature extraction stage.
• Encoding the normalized texture pattern uses two dimensional Gabor wavelets to convolve unwrapped image.
• 2D Gabor's mathematical model can result in real and values in outputs
• Normalized Hamming Distance quantifies iridal dissimilarity
• Masking code data removes noisy regions and disregards values
• The OR (⊕) detects discordant codes.
• The result from two perfect and similar codes will result in a 0.