Facial feature location using multilayer perceptrons and micro-features

An approach to robust feature location in images that treats the feature sought as a collection of micro-features is discussed. The spatial responses of multilayer perceptrons trained on micro-features are interpreted as probability distributions conditional on the image data. A postprocessor uses this information, together with prior information on the spatial relationships between micro-features, to choose the location of the feature that maximizes the a posteriori probability that the feature is at the given location. The method is demonstrated for the problem of locating the eyes in head-and-shoulders images, where it is shown to produce significantly better results than the use of a single detector trained to recognize the feature as a whole.<>