Boosting a haar-like feature set for face verification

This paper describe our ongoing work in the field of face verification. We propose a novel verification method based on a set of haar-like features which is optimized using AdaBoost. Seven different types of generic kernels constitute the starting base for the feature extraction process. The convolution of the different kernels with the face image, each varying in size and aspect-ratio, leeds to a high-dimensional feature space (270000 for an image of size 64x64). As the number of features quadruples the number of pixels in the original image we try to determine only the most discriminating features for the verification task. The selection of a few hundred of the most discriminative features is performed using the Ada-Boost training algorithm. Experimental results are presented on the M2VTS-database according to the Lausanne-Protocol, where we can show that a reliable verification system can be realized representing a face with only 200 features.