Gaussian weak classifiers based on co-occurring Haar-like features for face detection

摘要

Recently, in the context of appearance-based face detection, it has been shown by Mita et al. that weak classifiers based on co-occurring, or multiple, Haar-like features provide better speed-accuracy trade-off than the widely used Viola and Jones's weak classifiers, which use only a single Haar-like feature. In this paper, we extend Mita et al.'s work by proposing Gaussian weak classifiers that fuse information obtained from the co-occurring features at the feature level, and are potentially more discriminative. Experimental results, on the standard MIT+CMU test images, show that the face detectors built using Gaussian weak classifiers achieve up to 38 % more accuracy in terms of false positives and 42 % decrease in testing time when compared to the detectors built using Mita et al.'s weak classifiers.