CS328 Final Project Report

Algorithm

Basically, we only use histogram backprojection and histogram intersection( see M. J. Swain, D. H. Ballard: Color Indexing, International Journal of Computer Vision, Vol. 7, No. 1, 1991, p. 11-32 ) to do face detection. The algorithm could be divided into 3 stages:

Given several face templates, use histogram backprojection to locate possible facial regions in the given image. Several face templates are shown below:

We use Ben's and Li-Yi's faces for the training set, since from experiments we found that we can use Ben's face to find all white people, and use Li-Yi's face to find all east Asian people.
An example image with possible facial regions marked by black rectangles is shown below, click it for an enlarged image:
For each possible facial regions detected in the previous step, try to see if there is any eye in that region. We use histogram intersection of eye templates to do this. Several eye templates are shown below:

The image below shows the previous image with the non-eye regions knocked away is as follows:
Use histogram intersection to find the correlation between the facial templates and the possible facial regions survivied the previous step, and kick out those with correlation below some threshold.
The final image after kick out more regions is as follows:

Notice that we successfully detect the 2 faces, but also with one false match on Ben's left arm. That arm just looks too much like a human face in terms of histogram. We also found out that most of our false matches come from human arms.

liyiwei@graphics.stanford.edu