Adaptive and fixed eigenspace methods with a novel fitness measure for video based face recognition

In this paper a new system for identifying faces from video sequences using adaptive and fixed eigenspace approaches with a novel fitness measure is proposed. During the recognition process, each image in the gallery set is assigned a fitness value. The fitness value is updated for each frame and at the end of the probe video; the person corresponding to the gallery image with the highest fitness value is declared to be the identified person in the probe video. Eigenspace is used, for generating the feature vectors from the gallery set images. Two approaches have been introduced, where in the first approach; the eigenspace is updated after each frame is processed. The eigenspace update is performed by updating the fitness values and discarding the gallery images with the lowest respective fitness values. In the second method, a fixed eigenspace is generated from the initial gallery set and the fitness value for each gallery image is updated through the processing of the frames in the probe video. Again, in the end the gallery face image with the highest fitness value is declared to be the identified person. The BANCA video face database was adapted for performance testing. Both of the methods showed very competitive recognition rates in different scenarios.