Verification of faces in cluttered photographs

The human face is an object that is easily located in complex scenes by infants and adults alike. Our objective is to verify face locations hypothesized in photographs such as the ones typified by those found in newspapers. Our approach to face verification is based on a methodology of hierarchical rule-based system. The rules themselves are derived by a fuzzy model of assigning scores to the `goodness' of match between image features and model features. Scores of match computed with rules are further refined by a relaxation process. Face candidates are generated by a face locater (published elsewhere) and include face images as well as non-face images. Our objective is to confirm the face images while rejecting the non-face images. In an experiment on 80 face candidates returned by a face locater applied to newspaper photographs, the face verifier correctly identified all the faces as faces with a false positive rate of about 10%.