Face Recognition Algorithms Based On Principle Component Analysis And Independent Component Analysis

With the rapid development of technology, biology feature recognitiontechnology has become a convenient and e?ective authentication technology.Face is the most familiar model in human vision. The visual informationre?ected by face has important meaning and impact between people's inter-communion and intercourse. Because of its convenience, friendly and directadvantages, it has become a research problem with large academic and prac-tical values.The first chapter of this paper gives the definition of face recognition,introduces its general process and applications, compares the advantages anddisadvantages of several biology feature recognition technology, shows the di?-culties impacting recognition rate. This paper also introduces the technology'sthree development periods and their representative work. The first chapteralso introduces the main contents of face recognition technology, the recentstatus and several commonly used methods brie?y.Principal component analysis is a classical face recognition method. Thispaper gives its theory and the recognition process using eigenfaces method.After deeply researching on two-dimensional principal component analysis, wegive an improved two-dimensional principal component analysis.First we use 2DPCA on face image matrix X. Projecting X on U, we geta projection feature matrix Using Y's total scatter matrix, we get optimal projection axis U. Thestandard function is:Here, tr(SU) is the trace of SU, SU is Y's covariance matrix, thatDefinition of the image covariance matrix isEigenvectors (u1,u2,···up) corresponding to G's first p largest eigenval-ues composes matrix U which is the optimal projection matrix.We can prove that 2DPCA is equivalent to row-PCA. Following bilateral2DPCA, we do column-PCA, but di?erently we don't do column-PCA on X,we do it on Y.Similarly, we give Y's covariance matrix H, eigenvectors correspondingto H's first q largest eigenvalues composes matrix V which is the optimalprojection matrix. Then we get the last feature matrix ZT = Y TV . ThatIn the improved two-dimensional principal component analysis, U andV are related on theory, so images can be compressed better, recognitionrate will not drop. In addition, compared with PCA 2DPCA and bilateral2DPCA, it needs less factors, not only saves storage space, but also reducesthe feature extraction time. Therefore the improved two-dimensional principalcomponent analysis is superior. Independent component analysis is a recently developed face recognitionmethod based on high order statistical information, its features has good clas-sification ability and is independent with each other, it's not easily in?uencedby light. This paper introduces ICA's background-cocktail party problem,ICA's basic theory, and the pretreatment process with independent componentanalysis for face recognition. The paper also introduces an algorithm-FastICAwhich is e?cient and the convergence is fast. In view of the advantages ofWPCA, we use WPCA on ICA to reduce the dimension during pretreatmentprocess.Suppose the training samples of face images are X = (x1,x2,···,xm)1.Centering the training samples Here, (i,j)indicates the pixels'position in the image, (x,y)indicatesweighted center.4.Using FastICA to calculate separation matrix5.Classification and recognitionWe use YALE and ORL face database to do some experiments for twoimproved algorithms above. According to the numerical results, we can see:compared with PCA, 2DPCA and bilateral 2DPCA, the improved 2DPCAhas some advantages in recognition rate, extraction time, classification time;compared with PCA, WPCA and ICA, the algorithm combining WPCA withICA also have some improvement in recognition rate.Although the improved 2DPCA has some advantages in time and recog-nition rate, but it is similar to the PCA, just remove binary correlation. In ap-plications the images have some limitations. The algorithm combining WPCAwith ICA has advantages in recognition rate, but the calculation process ofFastICA is complicated, which causes this algorithm has disadvantages intime. This would require further improved research.