Research Of Face Recognition Pivotal Technique

In daily life, human's face is frequently used to identify people around and is the most normal pattern in one's vision. The vision information that faces reflect plays an important part of communications and contacts among people. As a result, human's face becomes the easiest acceptable identification method and one of the most potential ones.In this thesis, we first introduced the background and the main methods of the face recognition and then proposed a method of Eigenface based on the wavelet transform and fisher LDA algorithm for face recognition. Eigenface method for face recognition raised by M.Turk and A.Pentland is still the most popular face recognition algorithm. Although it is simple and effective, the method demands high unitary transform of images of the face inputed. So it is deeply affected by changes of illumination and postures. Therefore in this thesis, we first preprocessed the face images to eliminate the effects of illumination difference of them, and then the wavelet transform was used to obtain the stable low frequency sub-band of the face image in relatively low dimensions. In order to eliminate the correlations between entities of the image vector, the principal component analysis (PCA) and fisher LDA algorithm was used. High dimensions are projected into the Most Discriminating Space by Fisher algorithm, obtaining effects of extracting classifying information and compress characteristic space dimensions. Pattern samples have maximal distance between classes and minimal distance in classes in the new subspace after projection. Namely, pattern obtains the most separability in this space. The method proposed here significantly decreases the complexity of the algorithm and the characteristics extracted can have a better representation of the differences among different faces, which results in a high recognition and a high robustness of the algorithm. In this thesis, two kinds of face recognition methods are used, which have respective advantage and disadvantage. Experimental results presented in this thesis verified that the proposed algorithm is...