We have developed an automatic system for recognition and interactive search in the FERET face database. A recognition accuracy of 99.35% was obtained using two frontal views of 155 individuals. The figure below shows the result of a typical Photobook similarity search on the FERET database. The face at the upper left was selected by the user; the remainder of the faces are the most-similar faces from the 575 frontal views in the FERET database. Note that the first four images (in the top row) are all of the same individual (with/without glasses and different expressions). Also note this database represents a realistic application scenario where position, scale, lighting and background are not uniform. Consequently, our Automatic Face Processing System is used to correct for translation, scale, and contrast. Once the images are geometrically and photometrically normalized, they can be used in the standard eigenface technique.
NOTE: For information regarding the US Army FERET Database contact Jonathan Phillips.