Research On Face Image Feature Extraction And Classification Algorithm

Feature extraction and classifier design are the key problems in face recognition technology. Feature extraction is a kind of data map from a high-dimensional image to a low-dimensional space. It is very important to classification. Classifier is to separating extract characteristic with higher efficiency. Feature extraction has many classical algorithm, the famous based on subspace methods including PCA, LDA and Fisherfaces. Subspace algorithm is simple and effective, so widely applied in practice. Wavelet analysis is the famous data processing and analysis tool in signal processing field and also widely applied in pattern recognition. Neural network can realize any nonlinear mapping from input to output. So it has strong adaptability for processing high-dimensional image and is prominent in effect of classification recognition.This thesis summarizes the face feature extraction, classification and identification technology. Studies wavelet analysis combined with subspace methods and face classifier design technology.A kind of feature extraction and two classifier designs were put forward in this thesis:Firstly, wavelet transform algorithm suitable for face recognition based on detailed analysis of wavelet transform was proposed. This method is to distribute the wavelet transform coefficient weight, then combine subspace algorithm for feature extraction. This can improve the identification precision and time.Secondly, A kind of face classification algorithm was put forward based on RBF neural network optimization classification. This algorithm optimizes the radial basis function (RBF) network center vector through improved genetic algorithm. This can be representative to increase network approximation precision and still has high recognition efficiency in small training samples. Thirdly, based on learning vector quantization (LVQ) network design was proposed for face classifier design. This method analyses self-organizing neural network theory. The improved self-organizing neural network (learning vector quantization network) is applied for classification design based on analysis of self-organizing neural network theory. In small samples it can have better recognition rate and speed.The feasibility and effectiveness of these algorithms have been demonstrated through extensive experiments on several face databases.