Research And Implementation Of Face Recognition System Based On Linear Discriminant Analysis

Face recognition technology is a very popular research topic in the field of pattern recognition. It is used broadly in customs, public security, banking and other departments owing to its good properties which are direct, convenient, friendly and easily accepted by users. It involves image processing, pattern recognition, physiology, machine vision and other fields. General face recognition technology usually consists of face detection, image preprocessing, feature extraction and recognition techniques, face detection and feature extraction are critical components in a face recognition system.The thesis deeply analysis the principles of AdaBoost algorithm, Principal Component Analysis (PCA), Fisher Linear Discriminant Analysis (Fisher LDA, LDA) LDA method based on Maximum Scatter Difference(MSLDA) and Two-dimensional Linear Discriminant Analysis (2DLDA). Based on this research work, three innovative algorithm is presented as follows.(1) An autofocus human face detection method combining complexion detection with AdaBoost algorithm is proposed in this paper. In the new method, complexion information is used at first to locate the area of skin, then we calculate the number of pixels in the skin region and adjust the image according to the number of pixels, finally AdaBoost algorithm is adopted to detect human face. Experimental results show that this method reduces the time of face detection compared with Adaboost algorithm, the detection rate is increaced by 2.26% and false detection rate is reduced by 2.56% due to eliminating the possibility of face detection in non-skin region and improving the drawback of long-distance face detection.(2) A novel block LDA method based on half-even images(BHELDA) is proposed. Half-even images are used as samples in this method and each of these images is divided into several blocks, then extract sub-block features by LDA method, finally we integrate block features to do face recognition. The algorithm described above has been simulated on the platform of MATLAB and experimental results show that the method reduces the dimensions of samples, weakens the impacts of postures and facial expressions, takes the local features into account and it's recognition rate is up to 100%. (3) An improved 2DLDA (HE2DLDA) algorithm based on half-even images is proposed. The method extracts facial features directly from half-even images by using 2DLDA algorithm. Compared with LDA,2DLDA method not only solves the Small Sample Size problem of LDA algorithm, but also obtains a more accurate scatter matrix whose dimensions are greatly reduced, so it shortens the time of feature extraction considerably. Based on 2DLDA algorithm, half-even images are used as samples in the HE2DLDA algorithm and can further decrease the dimensions of scatter matrix, making it's feature extraction time more perfect.Finally, an entire face recognition system is developed on Visual C++ 6.0 and OPenCV1.0 platform and a satisfied recognition result is acquired. It realizes the value of face recognition technology in practical application.