An improved SNoW based classification technique for head-pose estimation and face detection

Abstract

A novel technique of reduction of the significance of overlapping features for efficient classification of complex patterns based on sparse network of windows where the features are the intensities at each pixel location of an image is proposed in this paper. Theoretical analysis performed on a set of patterns with overlapping features shows that the reduction of the significance of those features will improve the distinctiveness of the classifier. The methodology of classification is implemented in determining the pose and orientation of the face images in this paper. Classifying a face image of a particular pose from the rest of the face images with pose angles different from the first is essentially a two class problem. The probability distribution of the intensities at each pixel location over the entire training database of images is determined for both the classes and a measure of significance of the features is obtained based on the closeness in the relative probabilities of the two classes at that pixel. Features with equal probabilities are given least significance and features with largest difference in probabilities of the two classes are given highest significance. An efficient multilevel architecture for face detection with multiple classifiers for various face poses and orientations, keeping in view of the inherent symmetry of human face is also presented. The multiple levels in the classifier architecture deal with images of face regions in different degrees of orientations, poses and rotations in a hierarchical manner. An optimum image handling methodology resulted in reducing the number of classifiers required in the multilevel architecture to approximately half. Investigation of accuracy of head-pose estimation using the proposed technique is carried out. The proposed classification technique along with the architecture has been successful in discriminating face images whose pose angles are 100 apart. Comparison with other recent multiclass classification approaches in the context of pose estimation is carried out and it is observed that the technique is better both in terms of speed and accuracy