The basic outline of our aproach is as follows:

STEP 0: Getting the database
 

We have a 1250 .pgm pictures of 20 people and point files for every .pgm. Point file contains 20 coordinates  about important areas of the face, like the center of the eye, top of the nose etc.
 

STEP 1: Separation of 20 people
 

STEP 2: The Average maker

With each average we do a contrast equalizatio. This is very important because in this way we can make larger distances between the gray levels, therefore we have more information about the picture and also have a better result for matching.
 

STEP 3: Extracting key point coordinates from the database

We have 1020 point files about the separated 20 people and make a database containting on the first position the identifier, and then, the other 20 positions with the coordinates.
 

STEP 4: Normalizing the database
To compensate for slight rotation of the face, we scale all key points of a face such that the maximal x- and y- coordinate are 1, and the minimal ones are -1. This data can also be learned by a Neural Network more easily.

STEP 5: Matching points in input images

To locate key points in the face, we search small 64x64 patterns. Our Matlab implementation is based on normalized correlation of small image rectangles in the image space (this statistical similarity measure for image information compensates for illumination changes). A position of maximal correlation between the input image and the pattern is considered as a match.

Image-space based correlation computation has been used in stereo vision extensively and is very cost-intensive. A speedup from 1,5 minutes to 6 seconds per feature could be achieved using a hierarchical matching approach, where correlation is first computed at lower resolution levels, and the results are refined for elvels of higher resolution.

For an evaluation, see XXX.

For future improvements of our localization procedure, see YYY.

STEP 6: Classifier Input

We also validated that key point are suitable features for face recognition. Therefore, we extracted the coordinates of 20 key points for each image and concatenated them to a 40-dimensional vector. The coordinates were further normalized (see STEP 4)

STEP 7: Classification

As a classification, it emerged that a simple 1-Nearest-Neighbor approach was absolutely sufficient and performed satisfyingly in experimental results (see ZZZ).

We also implemented a Multilayer Perceptron approach in Matbab, but were not able to test it due to time pressure.