Feature Extraction Using Conformal Geometric Algebra for AdaBoost Algorithm Based In-plane Rotated Face Detection

Abstract

In this paper, we propose a novel face detection method based on the AdaBoost algorithm. In the past few years, a variety of variant AdaBoost approaches has been proposed and obtained increasing success in both performance and robustness. However, those approaches have not focused much on the geometric characteristics of face images. A new way to apply AdaBoost is introduced in this paper with the utilization of Conformal Geometric Algebra for extracting features from input samples. By analysis and experiments, using Conformal Geometric Algebra, we can find the hyper-spheres (-planes) that mostly fit data points and yield very low error for classification in both frontal and in-plane rotated face detection. Haar-like patterns are used as well but in a more pertinent approach to achieve more informative features. In comparison with the state-of-the-art face detection method developed by Paul Viola and Michael J. Jones, our proposed method can gain approximately the same accuracy as their framework. From a sufficient number of experiments, we prove that a small subset of the feature set can be totally adequate to develop a strong classifier with comparable performance and accuracy. Thus, iterating over the whole set of features to choose the best one for each round of AdaBoost is no longer needed. Besides, the memory usage for training is significantly reduced if we compare with the Viola-Jones system and a large amount of RAM is no longer required to store feature values as in their method.