Efficient face detection with multiscale sequential classification

The paper presents a sequential classification approach to improve the efficiency in visual object (face) detection. To reduce the computation while maintaining detection accuracy, a two-level hierarchy of sequential classification is proposed. At the top level, the overall detector is built on a cascade of classifiers at multiple resolution scales produced by a wavelet transform. Classifiers at low-resolution scales quickly rule out the regions likely to be background. Only object-like candidates are passed to subsequent high-resolution scales for more expensive tests. At the bottom level of the hierarchy, each classifier is implemented as a sequential Bayesian test using the features within the scale. The features are ranked adaptively according to their discrimination ability, which also leads to a quick decision. We demonstrate the scheme by an example of frontal view face detection.