Face Detection and Modeling for Recognition - Biometrics Research ...
Face Detection and Modeling for Recognition - Biometrics Research ...
Face Detection and Modeling for Recognition - Biometrics Research ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
size) to generate potential face c<strong>and</strong>idates, within the entire image. The algorithm<br />
first estimates <strong>and</strong> corrects the color bias based on a novel lighting compensation<br />
technique. The corrected red, green, <strong>and</strong> blue color components are first converted<br />
to the Y C b C r color space <strong>and</strong> then nonlinearly trans<strong>for</strong>med in this color space (see<br />
<strong>for</strong>mulae in Appendix A). The skin-tone pixels are detected using an elliptical skin<br />
model in the trans<strong>for</strong>med space. The parametric ellipse corresponds to contours of<br />
constant Mahalanobis distance under the assumption of the Gaussian distribution of<br />
skin tone color. The detected skin-tone pixels are iteratively segmented using local<br />
color variance into connected components which are then grouped into face c<strong>and</strong>idates<br />
based on both the spatial arrangement of these components (described in Appendix<br />
B) <strong>and</strong> the similarity of their color [84]. Figure 3.1 shows the input color image, color<br />
compensated image, skin regions, grouped skin regions, <strong>and</strong> face c<strong>and</strong>idates obtained<br />
from the face localization module. Each grouped skin region is assigned a pseudo<br />
color <strong>and</strong> each face c<strong>and</strong>idate is represented by a rectangle. Because multiple face<br />
c<strong>and</strong>idates (bounding rectangles) usually overlap, they can be fused based on the<br />
percentage of overlapping areas. However, in spite of this postprocessing there are<br />
still some false positives among face c<strong>and</strong>idates.<br />
It is inevitable that detected skin-tone regions will include some non-face regions<br />
whose color is similar to the skin-tone. The facial feature detection module rejects<br />
face c<strong>and</strong>idate regions that do not contain any facial features such as eyes, mouth,<br />
<strong>and</strong> face boundary. This module can detect multiple eye <strong>and</strong> mouth c<strong>and</strong>idates. A<br />
triangle is constructed from two eye c<strong>and</strong>idates <strong>and</strong> one mouth c<strong>and</strong>idate, <strong>and</strong> the<br />
best-fitting enclosing ellipse of the triangle is constructed to approximate the face<br />
59