| Summary: | 碩士 === 國立中正大學 === 資訊工程所 === 93 === In this thesis, we present a novel face authentication scheme by using appearance models and Hidden Markov Models. In our face authentication system, it can be roughly divided into two parts. First, the appearance model is used for features extraction, because an appearance model can not only extract the texture information, but also extract the shape information. We consider the shape information of a face is useful for the face authentication. Thus, we train an appearance model with a training set of labeled image sequences and then use this model to extract the low dimensional features of every image. In order to construct a face authentication system, we apply a vector quantization scheme to classify these features and combine the HMM to make full use of the temporal information across the video sequences. After all parameters in HMM are calculated, we can determine the thresholds dynamically for face authentication. An iterative algorithm with these thresholds is also proposed to select a suitable state number in HMM and a suitable class number of observations, because the performance of face authentication is affected by both variables.
As the result of experiment, we can show that our proposed video-based face authentication system works well on our constructed database. This database contains sixty-four video face sequences for training and sixty-four video face sequences for testing.
|
|---|
