An automatic face recognition system using the adaptive clustering network

III We have developed an Automatic Face Recognition (AFR) System that uses the Adaptive Clustering Network (ACN)-a hybrid classifier that combines neural network learning with statistical decision making. The ACN automatically groups similar faces into the same cluster and creates new clusters for novel input faces. During training, the ACN updates the clusters continuously, and multiple clusters are created for the same subject to accommodate variations in the presentation of the subject's face (for example, changes in facial expression andlor head orientation). With incremental training, new subjects and further variations of existing subjects may be added on line without retraining the classifier on previous data. During the testing process, the ACN associates an input face with the cluster that most closely matches the face. The ACN minimizes misidentifications by reporting completely novel input faces as "unknown." In addition to the ACN classifier, the overall APR system includes a preprocessing stage that removes the background in an image and centers the face in the image frame, and a feature-extraction stage that compresses the data. The system requires relatively simple processing and has been implemented in software on a SUN workstation. The preliminary results have been encouraging. Using imagery of eight subjects taken with a video camera, the system achieved a correct-classification performance of 99% with no misidentifications.