Automatic low complexity face detection algorithm

advertisement
Automatic low complexity face detection algorithm
Abstract
Transmission of video sequences over the standard phone lines is still a problem.
One solution is to make a non-uniform quantization of different parts of the image
depending on our interest in that part. Faces are usually an interesting object in a
video sequence.
In this project we have studied and implemented two different methods of face
detection. The first method was based on fitting a pattern and the second was
detection according to skin color.
Due to the elliptic shape of the human head it is proposed in the first method to
recognize the human face from it’s shape. This algorithm works on the edges image.
In order to disqualify other elliptic bodies each possible ellipse is checked to have an
axis ratio close to the one human face has, and a minimal degree of symmetry inside
it. Although the algorithm works, it prefers objects that are more elliptical then the
human head. Due to many search operations the algorithm is not a low complexity
one.
The second method proposed face detection according to skin color. The major
difference between skin color of different people is the luminance component, while
the chromatic components are almost the same for all humans. This algorithm works
on the Cr-Cb chrominance plane. A skin color statistic is generated and according to it
a binary image (face and non face) is created. Each region is separated by a bounding
rectangle from the others and is checked for the appropriate axis ratio and symmetry.
The algorithm gives good results, it also detects more than one person in an image and
at various face sizes.
This method has also relatively low complexity.
Download