Hand Geometry CSE 717 Why Hand Geometry l

Hand Geometry CSE 717

Why Hand Geometry l Acceptance Un. Intrusive l Ease of Collection Of Data l Fast capture/processing l Small template size l Uses include Access control, time & attendance, border crossing

Feature Extraction l Image Capture l Preprocessing l Measurements l Optimization of the template Size l Feature Selection & Feature Vector Size

Image Acquisition Using Pegs Free Hand geometry

Image Capture l Flat Bed Scanner 150 dpi scanner. l CCD color camera. Color Photograph in the form of a Jpeg format. l The Lateral view of the Hand can also be captured by the mirror placed in the side to measure heights.

Using Pegs Raul Sanchez-Reillo Carmen Sanchez-Avila Ana Gonazalez-Marcos

Preprocessing l Binarize the Image. l Rotation n Resizing. l Extract the Contour of the Image.

Preprocessing l 1 st step: Binarize the Image. I (bw) = [(Ir + Ig) – Ib] l Resizing n Rotation. Deviation of the hand are corrected l Edge Detection Algorithm eg. Sobel Function

Problems with Using Pegs Alexandra L. N. Wong 1 and Pengcheng Shi 2 Deformation of the Shape of the Hand by the Pegs. Different Placements of the Same Hand

Landmark Extraction Hand Alignment Applying the border-following Algorithm 1 Alexandra L. N. Wong 1 and Pengcheng Shi 2 Application of the border-following Algorithm

Feature Extraction Lengths of four fingers Widths of four fingers at 2 locations Shapes of the fingertips Alexandra L. N. Wong 1 and Pengcheng Shi 2

Hierarchical Recognition l Class I : 13 finger lengths and the finger widths Gaussian mixture Model is used to classify these features Andrew W. Moore Associate Professor School of Computer Science Carnegie Mellon University l Fingertips - class II www. cs. cmu. edu/~awm

Gaussian Mixture Modeling l Approach bet Statistical n Neural Networks l Modeling the patterns with determined number of Gaussian Models l Weighing Coefficient of Gaussian Model Mean Covariance Vector are the characteristic parameters.

GMM cont’d l Preset Threshold Value of the GMM Probability Estimation. l Group II features - Euclidian Dist Measure bet the sample template & the given template. Threshold is used to reject the templates. (eg. 2 pixels)

Results l Hit Rate : Typical Methods of Comparisons Euclidean Distance Measure Group 1 and 2 Hit Rate 1 0. 8889 FAR 0. 1222 0. 022 Alexandra L. N. Wong 1 and Pengcheng Shi 2

Alexandra L. N. Wong 1 and Pengcheng Shi 2 GMM threshold’s role in reducing the FRR.

Different Comparison Algorithms

Work Done by Other Researchers l Raul Sanchez-Reillo Carmen Sanchez-Avila Ana Gonazalez-Marcos have done the development of the GMM based comparison Algorithms. l Alexandra L. N. Wong 1 and Pengcheng Shi 2 : Pegs Free Hand based Geometry

Observations: l GMM obtain the best results l the other possible comparison algorithms are Euclidean Hamming Distance based , Radial Basis Function RBF Neural Networks. l GMM based template require much more memory than the other comparison based templates.

Conclusion l Ideal Future Research for Medium and Low Security based Biometrics. l Can be used together with other Biometrics Palm Prints l Non Geometrical hand features such as color can be used.