Automatic location of visual features by a system of multilayered perceptrons

A two-stage neural vision system for locating facial features is described. The first stage generates search regions in an image at low spatial resolution, and the second pinpoints the features at high resolution. Both stages employ multilayered perceptrons trained to detect specific visual details, followed by sophisticated global postprocessing of their outputs. This work demonstrates the power of combining neural feature detection with knowledge-based context-sensitive methods.