Biomedical Person Identification via Eye Printing

transcript

Biomedical Person Biomedical Person Identification via Eye Identification via Eye

PrintingPrinting

Masoud Alipour Masoud Alipour (malipour@ipm.ir)(malipour@ipm.ir)

Ali Farhadi Ali Farhadi (farhadi@ipm.ir)(farhadi@ipm.ir)

Nima Razavi Nima Razavi (n_razavi@ce.sharif.edu)(n_razavi@ce.sharif.edu)

IPM – Scientific Computing CenterIPM – Scientific Computing CenterVision GroupVision Group

Institute for Studies in Theoretical Physics and MathematicsInstitute for Studies in Theoretical Physics and MathematicsTehran-IranTehran-Iran

OutlineOutline

Introduction to human eye and Iris structureIntroduction to human eye and Iris structure Human Eye and Iris structure and properties of Human IrisHuman Eye and Iris structure and properties of Human Iris

Image De-noising Image De-noising Application of wavelet analysis.Application of wavelet analysis.

Iris LocatingIris Locating Creating Edge Image and Circular Hough Transform.Creating Edge Image and Circular Hough Transform. Find Ciliary and Pupillary Boundaries. Find Ciliary and Pupillary Boundaries.

Feature Extraction Feature Extraction Application of Higher Order Statistics (creating LPC Matrix).Application of Higher Order Statistics (creating LPC Matrix). Discrete Cosine Transform (DCT).Discrete Cosine Transform (DCT). Analysis of Geometric Characteristics of obtained Surface.Analysis of Geometric Characteristics of obtained Surface. Frequency Domain Analysis and FFT.Frequency Domain Analysis and FFT.

Feature ClassificationFeature Classification

Introduction to human eye Introduction to human eye structurestructure

Eye Structure :Eye Structure :

Fig1.Human Eye Fig1.Human Eye

Human Iris StructureHuman Iris Structure

Anterior layer of Human Iris :Anterior layer of Human Iris :

1. Pigment frill1. Pigment frill

2. Pupillary area2. Pupillary area

3. Collarette3. Collarette

4. Ciliary area4. Ciliary area

5. Crypts5. Crypts

6. Pigment spot6. Pigment spot

Biometric Properties Of Human IrisBiometric Properties Of Human Iris

featuresfeatures

• 1. crypts . 1. crypts . • 2. pigment spot.2. pigment spot.• 3. radial and concentric 3. radial and concentric

furrows .furrows .• 4. collarette.4. collarette.• 5. pigment frill.5. pigment frill.

Concentric Concentric furrowsfurrows

CollaretteCollarette

Radial furrowsRadial furrows

OutlineOutline

Feature Extraction Feature Extraction Application of Higher Order Statistics.Application of Higher Order Statistics. Discrete Cosine Transform (DCT).Discrete Cosine Transform (DCT). Analysis of Geometric Characteristics of obtained Surface.Analysis of Geometric Characteristics of obtained Surface. Frequency Domain Analysis and FFT.Frequency Domain Analysis and FFT.

Image De-noising Image De-noising

Application of Daubechies wavelet to removeApplication of Daubechies wavelet to remove

1.1. High frequency noise introduced by High frequency noise introduced by cameracamera

2.2. Reflection noiseReflection noise

OutlineOutline

Feature ClassificationFeature Classification Neural Networks for classificationNeural Networks for classification

Iris Locating Iris Locating

Iris Locating is achieved by :Iris Locating is achieved by :

Creating Edge-ImageCreating Edge-Image Circular Hough Transform of Edge Image.Circular Hough Transform of Edge Image. Locating Ciliary Boundary.Locating Ciliary Boundary. Locating Pupillary Boundary . Locating Pupillary Boundary . Creating Iris Image ( Polar indices ).Creating Iris Image ( Polar indices ).

Circular Hough TransformCircular Hough Transform

1. Description of 1. Description of circular Hough circular Hough spacespace

2. Normalizing the 2. Normalizing the Hough SpaceHough Space

3. Locating center 3. Locating center and radius of the and radius of the cilirary boundary.cilirary boundary.

(x,y)(x,y)

Iris LocatingIris Locating

Results :Results :

Fig Fig 1.1.

Fig 2.Fig 2.

Original Image Original Image Edge-ImageEdge-Image

Iris LocatingIris Locating

20 40 60 80 100 120 140 160

Fig 3. Fig 3. Maximum pointMaximum point

Fig 4.Fig 4.

Circular Hough SpaceCircular Hough Space

(one layer) (one layer)

Iris ImageIris Image

OutlineOutline

Feature ExtractionFeature Extraction

-Application of Higher Order Statistics.-Application of Higher Order Statistics. -Discrete Cosine Transform (DCT) Analysis.-Discrete Cosine Transform (DCT) Analysis. -Analysis of Geometric Characteristics of Surface -Analysis of Geometric Characteristics of Surface

of LPC coefficients.of LPC coefficients. -Frequency Domain Analysis and FFT.-Frequency Domain Analysis and FFT. - Circular DCT - Circular DCT

Higher Order StatisticsHigher Order Statistics Creating SectorsCreating Sectors

1.1. Each sector is defined by 4 Each sector is defined by 4 parameters (rparameters (rminmin ,r ,rmax max ,th,thmin min ,th,thmax max ))

2.2. We create sectors from rWe create sectors from rmin min to rto rmax max

and moving counter-clockwise from and moving counter-clockwise from ththminmin to th to thmaxmax with large overlaps. with large overlaps.

overlapping Sectorsoverlapping Sectors

Higher Order StatisticsHigher Order Statistics Definition of LPC CoefficientsDefinition of LPC Coefficients

Neighborhood ConfigurationNeighborhood Configuration

SSS AAA 2021 ,...,,

Higher Order StatisticsHigher Order Statistics Definition of LPC CoefficientsDefinition of LPC Coefficients

Linear Predictive CodingLinear Predictive Coding

qYqXMinimize

qpXaqY

S = Sector IndexN = neighborhood configuration (o NN )X(p) = brightness of pixel p (value of the pixel)

SS AAA ,...,, 21

DCT Analysis DCT Analysis

1.1. From the average of the nearest four horizontal and From the average of the nearest four horizontal and vertical neighbors we obtain a matrix A. For ease of vertical neighbors we obtain a matrix A. For ease of references we call this matrix as PLPC.references we call this matrix as PLPC.

2.2. Defining a square w * w window W on the PLPC Matrix.Defining a square w * w window W on the PLPC Matrix.3. 3. Computing DCT Coef of each window.Computing DCT Coef of each window.4.4. As window W moves along a row , the curve C is obtained As window W moves along a row , the curve C is obtained

by calculating by calculating ||Differences of DCT coefficients of two contiguous windows ||||Differences of DCT coefficients of two contiguous windows ||2 2 5.5. Hence for each row we obtain a curve. Averaging these Hence for each row we obtain a curve. Averaging these

curves over different rows , we obtain a curve which we curves over different rows , we obtain a curve which we call FC. call FC.

6.6. Curve FC is the first part of our feature vector.Curve FC is the first part of our feature vector.

Feature VectorFeature Vector

A1A1 A2A2 A3A3 A4A4 A5A5 A6A6 A7A7 A8A8 A9A9 BB C C DD EE FF GG

DCT of PLPCDCT of PLPC

Geometric Characteristics of PLPC Surface Geometric Characteristics of PLPC Surface

Each sector is identified by Each sector is identified by

initial initial ρρ and and θθ . .

Each (Each (ρρ,,θθ ) together with ) together with corresponding entry of PLPC matrix corresponding entry of PLPC matrix give a surface (PLPC surface).give a surface (PLPC surface).

PLPC SurfacePLPC Surface

ZZs’s’

),,( ''' sss Z

),,( sss Z

),( ss

),( '' ss

Geometric Characteristics of PLPC Surface Geometric Characteristics of PLPC Surface

1.Trinagulation of PLPC Surface.1.Trinagulation of PLPC Surface. 2. Mapping gravity center of each triangle on plate 2. Mapping gravity center of each triangle on plate

z=0z=0 3. Centroid Matrix3. Centroid Matrix 4. Statistical invariants of Centroid matrix are next 4. Statistical invariants of Centroid matrix are next

elements of the feature vector.elements of the feature vector.

Triangulation

66333333

33115544

22336644

Centroid MatrixCentroid Matrix

Statistical invariants of Centroid matrixStatistical invariants of Centroid matrix

We make use of Mean , Variance and We make use of Mean , Variance and Kurtosis of Centroid Matrix.Kurtosis of Centroid Matrix.

These three invariants are next 3 These three invariants are next 3 elements of the feature vector.elements of the feature vector.

Recall that Recall that

Kurtosis(X) =E[XKurtosis(X) =E[X44] – ] – 3*E[X3*E[X22]]2 2 ..

Mean of Centroid Mean of Centroid MatrixMatrix

Variance of Centroid MatrixVariance of Centroid Matrix

Kurtosis of Centroid Kurtosis of Centroid MatrixMatrix

Frequency Domain Analysis and FFTFrequency Domain Analysis and FFT

1.1. Let D be the differences of consecutive columns in matrix Let D be the differences of consecutive columns in matrix of LPC Coef.of LPC Coef.

2.2. These quantities can be regarded as function on set of 20 These quantities can be regarded as function on set of 20 points.points.

3.3. Calculate FFT of this function. Thus transferring data to Calculate FFT of this function. Thus transferring data to Frequency Domain. (resulted in CFrequency Domain. (resulted in C2020))

4.4. Make use of absolute values to transfer data to RMake use of absolute values to transfer data to R2020..

5.5. Projecting the data to 3D subspace.Projecting the data to 3D subspace.

6.6. Application of Geometric Properties of 3d obtained scatter Application of Geometric Properties of 3d obtained scatter plotsplots

Geometric Properties of 3D scatter plotsGeometric Properties of 3D scatter plots

The next member of The feature The next member of The feature vector is the volume of the convex vector is the volume of the convex closure of the projected data.closure of the projected data.

Volume of the convex closure of fft Volume of the convex closure of fft coefcoef

Circular DCTCircular DCT

1.1. Scanning the Iris Layer by Layer( Each Scanning the Iris Layer by Layer( Each Layer is a circle ) and obtain Vector C. Layer is a circle ) and obtain Vector C.

2.2. Calculating DCT coefficients of C .Calculating DCT coefficients of C .3.3. By merging results of all layers, we obtain a By merging results of all layers, we obtain a

Matrix.Matrix.4.4. Kurtosis of this matrix is the next element Kurtosis of this matrix is the next element

of the Feature Vector.of the Feature Vector.

Kurtosis of circular DCTKurtosis of circular DCT

Analysis of geometric Analysis of geometric characteristics of CDCTcharacteristics of CDCT

1.1. Applying Circular DCT , we obtain a high dimensional data Applying Circular DCT , we obtain a high dimensional data set.set.

2.2. Make use of projection to reduce dimensionality of the Make use of projection to reduce dimensionality of the data to data to

1D data (by average)1D data (by average)

3. Triangulation of PLPC Surface.3. Triangulation of PLPC Surface.

4.4. Mapping mass center of each triangle on plate z=0Mapping mass center of each triangle on plate z=0

5.5. Centroid MatrixCentroid Matrix

6.6. Kurtosis of centroid matrix is the last element of the Kurtosis of centroid matrix is the last element of the feature vector.feature vector.

Kurtosis of circular DCTKurtosis of circular DCT

Kurtosis of Centroid Matrix of circular Kurtosis of Centroid Matrix of circular DCTDCT

Feature vector has been tested on a Feature vector has been tested on a small data base of about 35 irises.small data base of about 35 irises.

So far has produced no type 1 or type So far has produced no type 1 or type 2 errors.2 errors.

Remains to be tested on a large data Remains to be tested on a large data base.base.

Questions?Questions?

Biomedical Person Identification via Eye Printing

Documents