Font Size: a A A

Face Recognition Algorithm Based On The Positive Side Of The Cascade Matching Study

Posted on:2008-05-04Degree:MasterType:Thesis
Country:ChinaCandidate:S J ShiFull Text:PDF
GTID:2208360218450209Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Face recognition has become one of active research areas in pattern recognition and computer vision. It is going to a practical stage in constrained environments, after a research history of more than thirty years. However, the challenges that recognition algorithms faced include unconstrained environments (especially different lighting conditions,head pose),large-scale face database and so on.A frontal and lateral hierarchical matching algorithm is proposed in this thesis, to solve the problem of face recognition in large-scale face database. This face recognition algorithm is designed as hierarchical process. In the first matching, fast recognition is executed in large-scale face database and picks out several similar faces as the new face database. In the second matching, accurate recognition is executed and finds the right face. The main contributions in this thesis are as follows:(1)The choice of Gabor filter parameters for facial feature extraction is discussed, which is validated by experiments. The choice of kernel function for face recognition is discussed. And the parameters of Gaussian radial basis function in KPCA are validated by experiments.(2)At the stage of frontal face recognition, a recognition algorithm based on low frequency subband is proposed, which is combining Gabor feature and kernel principal component analysis. First, face image is decomposed by two dimension wavelet transform. Then the low frequency subband is convolved with Gabor filters and downsampled by a factor. After that, the kernel principal component analysis is used to reduce the dimension of Gabor features. Finally, minimum distance classifier is used to classify. Because of Gabor filters, this algorithm is more robust against variations in illumination and expressions. The experiment result in FERET face database shows that this method is better than KPCA and costs less time than not use wavelet decomposition.(3)At the stage of lateral face recognition, a matching algorithm based on nose profile curve is proposed. First, the lateral face image is pre-processed and the nose is partitioned. Then the continuous nose curve is obtained by thinning and removing the edge noise. After that the contour is produced by curve fitting and the curvature of every point of the contour is calculated. Finally, the similarity between two curvature lines is calculated by correlation function. And several similar faces are selected by threshold, which is validated by experiments. The experiment result in FERET face database shows that this method can be executed in large-scale face database quickly, and new face database which contains right face is generated.The experiment result shows that frontal and lateral hierarchical matching algorithm can shorten recognition time efficiently, and solve the problem of recognition time increased dramatically when the scale of face database increased.
Keywords/Search Tags:face recognition, Gabor filter, kernel principal component analysis (KPCA), lateral contour, curvature
PDF Full Text Request
Related items