This thesis proposes a real-time viewpoint invariant face detection and recognition system. This is motivated by the trends of face recognition research and the results of psychophysics and neuroscience.
So far the face recognition has been researched for the frontal face, but recently the viewpoint invariant face recognition has been researching. The reason is that the face recognition from previously unseen viewpoints is inherently more difficult than matching face at the same view.
And there is two biologically plausible hypothesis of human visual system. The first one is view-based representation, which provides features of facial images under the circumstances of 3D viewpoint rotation. The second hypothesis is temporal associative learning of invariance, which enables predictive generalization from one viewpoint image.
The proposed system is composed of three component modules which are face detection, normalization and recognition. First the face detection is based on integral image, Haar-like wavelet feature and AdaBoost algorithm. Second the face normalization is based on histogram equalization, edge detection and statistical standardization. Finally the face recognition is based on ISA basis representation, view classification and frontal face prediction.
The system was experimented on 15 subjects and showed 2-4 frame rates of face detection and recognition. The experiments showed high performance for frontal face but low for the rotated face.