We detect the face of the subject within a source image containing a single subject on a natural background. The face is in an upright position.
The detection of the head is based on the Y-Phase attentional operator. Y-Phase is a novel operator for the detection of smooth three dimensional convex and concave regions in an image. It is a reduced variant of a more general operator called D-Phase, which is currently employed for the detection of areas of interest in images.
The detection of the head enables the detection of the eyes by the radial symmetry transform . This transform detects areas exhibiting generalized symmetry in various direction.
Once the eyes are detected, one is able to transfer the eyes into pre-determined locations in an image of pre-defined size using an affine transformation. We call the image created by this process: the normalized image. The normalized image is robust to translation, rotation and scaling. This allows us to recognize subjects based on their normalized images, since the variance of the normalized images of the same subject is much lower than the variance of the original images of that subject.
The following animation demonstrates the normalization process. The steps shown in the movie are summarized below:
Original image containing a face.
Lighting improved by histogram equalization and logarithm.
Gradient magnitude.
Gradient argument.
Y-Phase.
Head detection by cross correlation marked on Y-Phase map.
Symmetry magnitude of head image.
Radial symmetry of head image.
2% histogram threshold applied to radial symmetry map.
Eyes marked on head image.
Eyes marked on original image.
The normalized image.