Face detection by generating and selecting features based on Kullback-Leibler divergence

Face detection from images is a complex and nonlinear problem due to the various kinds of face images. This problem is solved by conversion of the original feature vectors extracted from images into high-dimension feature vectors using nonlinear mapping, and then finding face/nonface discriminant functions in the mapping space. If such discriminant functions are based on the inner products of high-dimension vectors, such inner products can be easily obtained by substitute calculations of kernel functions in the original feature space. However, in conventional recognition algorithms using kernel functions, numerous features are required to improve recognition accuracy. This paper proposes a new face detection method that uses generation and selection of features on the basis of Kullback-Leibler divergence (KLD). KLD refers to a distance between the distributions of face and nonface data for certain features. Features with large KLD are used for face detection. Moreover, by evaluating the features based on their KLDs, we can generate new features, and deal with different kinds of features concurrently. In experiments, a classifier designed by the proposed method achieved high recognition performance, while using few features.