Online Face Recognition System

An online face recognition system automatically detects face regions, represents facial from the video, and compares the features with that in face database to recognize facial identity if a face is present. The target of this thesis is to build an online face recognition system. Accounting for the speed and recognition rates of the system, some research on face detection, face representation and face models extraction will be done in this thesis. The main work of the dissertation are summarized as follows:Color and cascade based face detection: Some studies have suggested that color does play a important role in face detection. When the color cue is available, one can reduce the search regions by pre-processing the images and selecting only the skin-like areas. Furthermore, color are the computational efficiency and robustness against some geometric changes such as rotation, scaling. In this dissertation, color cue firstly be used to pre-process the images and segment the skin-like areas, then, a cascade based face detection is presented to detect face in these skin-like areas.Face models extraction based on locally linear embedding algorithm: The key aspect of the face recognition algorithm is use of only a small amount of data(the most representative samples)for recognition, which should resume well the intra-class variability in face appearances caused by different poses and illumination changes, thus leading to low memory requirements and high speed processing, however, finding these representative samples or exemplars is not an easy task. In this dissertation, LLE algorithm is used to mapping the data onto a low-dimensional embedding space, then K-means clustering is performed. The set of exemplars is then defined as the cluster centers. This strategy is motivated by the demonstrated simplicity and efficiency of LLE to recover meaningful low-dimensional structures hidden in complex and high dimensional data such as face images.A new face representation algorithm based on Discrete Cosine Transform: As we know DCT is near optimal best and approaches the KLT, which is an optimal transform and successfully has beenused in face recognition. Thus, it is expected that it too will exhibit desirable pattern recognition capabilities. If this is shown to be the case, then the use of the DCT in face recognition becomes of more value than the KLT because of it's computational speed. Furthermore, fast algorithms for computing the DCT have been presented and DCT is independent of other signals change. In this dissertation, a new face representation algorithm based on weighted DCT fusing global facial feature and local facial feature is proposed. An online face recognition system: Implementing an online face recognition system by integrating the above algorithms into one frame, which can be used to develop and test new face recognition algorithms.