A method for eye detection based on SVD transform Somayeh Danafar Lila Taghavi Alireza Tavakoli
Outline The advantages of algorithm SVD in a nutshell The methodology Results from a number of images The result of interest points algorithm The algorithem’s errors
:The algorithm is robust relative to Changes in lighting Eye color and complexion Blurring Introduction of glasses Different resolutions Complex background Variability in scale and orientation
Singular Value Decomposition in a Nutshell U and V are orthogonal matrices i = i ( A T A) i=1,…,N
Different coefficients of Da f1(D a )= N-3 + N-2 + N-1 f2(D a )= 1 + 2 + 3
Eye Detection The algorithm proceeds in four steps : 1. Variance reduction 2. Application of SVD transform with a nonlinear function f 3. Application of edge detection algorithm 4. Separation of boundary edges f= 1/(1+s(t)) for t=1,..,8
Edge Detection f2(D a )= 1 + 2 + 3 g(D a )=f2(D a )f
Original Image SVD transform uses the exponential of a linear function of diagonal part of the SVD decomposition. SVD Transform in diffident coefficients of the sigma in the SVD Using Edge Detection Using noise removal -= minusequal The Methodology
- = SVD Transform Edge detection Edge detection with noise removal
Rotate :65 ْ
Rotate: 45 ْ
Variation in lighting
Introduction of glasses
Effect of change in orientation and closure of eyes
Application to images with a complex background
The SVD transform
The SVD transformsFinal result
The result of svd transform Final result Original image
SVD versus Interest Points algorithm 25 points 60 points 43 points 3 points 14 points 10 points
The errors in SVD: 1) Missing one eye 2) Detecting additional points