A Regularized Approach to Feature Selection for Face Detection

Abstract

In this paper we present a trainable method for selecting features from an overcomplete dictionary of measurements. The starting point is a thresholded version of the Landweber algorithm for providing a sparse solution to a linear system of equations. We consider the problem of face detection and adopt rectangular features as an initial representation for allowing straightforward comparisons with existing techniques. For computational efficiency and memory requirements, instead of implementing the full optimization scheme on tenths of thousands of features, we propose to first solve a number of smaller size optimization problems obtained by randomly sub-sampling the feature vector, and then recombining the selected features. The obtained set is still highly redundant, so we further apply feature selection. The final feature selection system is an efficient two-stages architecture. Experimental results of an optimized version of the method on face images and image sequences indicate that this method is a serious competitor of other feature selection schemes recently popularized in computer vision for dealing with problems of real time object detection.