Face Recognition Based On Subspace And Transformed Domain

Human faces are our primary focus of attention in social intercourse, playing a major role in conveying identity and emotion. Face recognition is the hotspot of research in recent years and it aims at endowing computers with the ability to identify different human beings according to their face images. Face recognition is closely related to many disciplines, such as computer vision, image processing, pattern recognition, artificial intelligence and cognitive psychology, et al, and face images involve many varying factors, so face recognition is a challenging topic of research. A great number of key issues are to be solved which mainly include: (1) the preprocessing of face images, such as face righting, image enhancing, normalizing and denoising, et al. (2) complete feature extracting, involving local and global feature, by geometrical and statistical methods. (3) high accuracy face classifiers, including classifiers based on distance and based on statistics.This thesis is under the background of face recognition and based on the study on one dimensional and two dimensional linear feature subspace and image transform methods, such as wavelet transform, discrete cosine transform (DCT) and Radon transform. The research work of the thesis focuses on image preprocessing, facial feature extraction and face recognition algorithm. Some specific problems are discussed and some improved algorithms are presented in the thesis.The main research work and innovations are as follows:1) Give a thorough survey of the face recognition technology. Firstly, this thesis introduces the content, significance and general situation at home and abroad. Then primary face recognition methods are categorized according to the form of facial feature. The classifying rules and public face image databases in common use are also listed. Finally, the technical trends in face recognition field are discussed.2) A median filter-based Laplacian edge detector is proposed. The edge is an important feature of faces. Facial edge detecting could be used in the facial outline extracting and region segmentation, et al. Conventional algorithm often gets false edge when detecting images with noises. In the thesis, based on the Laplacian operator, a model for edge detecting which consists of a ring filter and a noise-smoothing filter is first introduced. The Laplacian operator can not smooth noises because its noise-smoothing filter is indeed an all-pass filter, while median filter could smooth noises and protect edges in images. So a new edge detector is derived by substituting the median filter for the all-pass filter in Laplacian operator. An optimal threshold for edge detection is also presented. Examples show that the proposed edge detector has better ability to detect edges and smooth noises.3) A 3-channel orthogonal perfect reconstruction filter bank and an image denoising algorithm based on it are proposed. The filter bank is designed based on the theory of multi-phase representation. On the basis of Haar wavelet, its high-pass filters are capable of edge detecting; as a mean filter, its low-pass filter could smooth noise. Thus the filter bank is suitable for image denoising. Consequently, an image denoising algorithm using a moving window, soft threshold and the filter bank is deduced. The advantage of denoising with a moving window is that it can avoid the 'mosaic' effect of the filtered image. In addition, on account of a pixel in an image bearing no relation to the pixels far from it in distance, a small size moving window will produce good result. The algorithm fully makes use of the correlation properties of the two groups of wavelet transform coefficients and gets the optimal estimates of coefficients, so it is effective on image denoising.4) Through the study on Linear Discriminant Analysis (LDA) and its improved algorithms, a new feature extraction method called weighted LDA-based null space is proposed and the selection of weight parameters is discussed also. The method weights the eigenvectors of the non-null space of between-class scatter matrix in direct proportion to the corresponding eigenvalues. The eigenvectors in correspondence with bigger eigenvalues overweigh the eigenvectors with smaller eigenvalues when evaluating the optimal projection directions. It not only avoids the small-sample-size problem, but also is easier to divide the different classes after the data is projected.5) An efficient face recognition method combining the proposed 3-channel orthogonal filter bank and two-dimensional LDA (2DLDA) are presented based on the study on two-dimensional linear feature subspace and Wavelet Transform. By this method, each face image is first decomposed into nine subimages by use of the 3-channel filter bank at the first level. Then 2DLDA is used to extract the feature from each subimage and the distance of each subimage between a training sample and a test sample is get. The final distance between a training sample and a test sample is acquired by weighting and fusing the distance of each subimage. Experimental results show that the proposed method is better than the original LDA and 2DLDA in terms of recognition accuracy.6) A new feature extraction method (2DDM) based on DCT and two-dimensional maximum margin criterion (2DMMC) is presented. We also prove that 2DMMC can be directly implemented in the DCT domain and the classification results by the measure of Euclidean distance are exactly the same as the one obtained from the spatial domain. For images compressed using the DCT, 2DMMC can be directly implemented such that the inverse DCT can be skipped and the computational cost is reduced. The experimental results show that 2DMMC achieves better face recognition performance than 2DPCA and 2DLDA, further 2DDM acquires higher recognition rate and is less time consuming than 2DMMC when appropriate DCT coefficients are retained.7) After the study on the theory of Radon transform and Finite Radon transform (FRAT), a novel feature extraction method using FRAT is proposed. Radon transform is a powerful tool to capture the directional information in images, but its transform accuracy depends on the amount of projection directions and it is difficult to reconstruct the original image completely. Whereas FRAT does not have the disadvantages and can be calculated conveniently. There are lots of curve information of facial outline and features in a face image. If an image is divided into blocks, the curve edge or texture in each block could be viewed as straight lines approximately, and then FRAT could be used to each block. By the method, each face image is first projected into FRAT domain by FRAT; then main FRAT coefficients are selected by soft threshold; finally, 2DMMC is used to extract features for face classification from the main FRAT coefficients. The experimental results on face databases show that the proposed method is feasible and efficient.