Improved Algorithm Of LDA And Its Application In Face Recognition

With the rapid development of computer science and Internet technology, biometric technology is widely used in our social life, such as fingerprint recognition, face recognition and so on. As the face recognition has non mandatory, non contact and so on, and, as a biometric technology, face recognition method quickly become an important research area.Face recognition is an identification based on the method of human face feature information achieved. At present, researchers proposed many different face recognition methods, and developed different kinds of face recognition system. However, in the practical application, the face recognition method or system is exposed a lot of problems, which are mainly from two aspects: first, face images are easily affected by external environment influence, such as rain, snow, wind, mud and others. The existence of these factors will cause different degree of noise interference in face detection and recognition process. Second, the dimension of the face image data is relatively high, which often are needed to reduce the dimension of the high dimensional data in the practical application. The different types of data dimensionality reduction algorithm will affect the feature extraction efficiency, thereby reducing the accuracy of face detection and recognition.In view of the above problems, LDA feature extraction methods and the noise of image processing were studied in this paper and a total scatter matrix of null space LDA algorithm is obtained. The method is applied to the problem of face recognition system. The specific research work mainly includes the following contents:1. Image denoising. Consider to complex environmental factors influence outside, in particular, the noise interference, through the analysis of the respective advantage of salt and pepper noise characteristics and a variety of filter, with salt and pepper noise image, using median filter to remove the noise, combined with ORL and YALE face database on the experimental study, the paper discusses the influence of noise density on the parameter setting of feature extraction algorithm.2. The feature extraction of LDA and its extended method is analyzed and studied. Through the analysis of the traditional LDA feature extraction method, and the use of fisher criterion, the methods of LDA based feature extraction are studied, mainly including combining PCA(principal component analysis) and LDA feature extraction method, direct LDA method, null space LDA, orthogonal LDA method, uncorrelated LDA method, regularized LDA method, kernel LDA method, tensor LDA method and LDA method based on two-dimensional image. By choosing the nearest neighbor and k-nearest neighbor classifier, the experiments of different feature extraction methods in face recognition performance are studied.3. For the null space LDA Algorithm defects, and based on total scatter zero space feature extraction, it put forward the total scatter of null space LDA Algorithm. At the same time, by introducing dynamic weight adjustment within-class scatter matrix and between-class scatter matrix providing classification information, in order to adapt to the effects of different levels of noise. Combined with the face recognition problem, it chooses the nearest neighbor classifier and k-nearest neighbor classifier experiments to study the performance of face recognition. Finally, based on the proposed feature extraction algorithm, the face recognition system is designed and developed.