Research On Methods Of Subspace Feature Extraction In Face Recognition

Face recognition is a hot research field in biometrics identification technology, andhas a very important research significance and practical value. A complete facerecognition system consists of four main parts: face detection, pretreatment, facialfeature extraction and separation of classes. Among, Extracting effectivelydiscriminated features is a critical factor for face recognition and affects theperformance of the whole face recognition system. And among all the proposed featureextraction methods, subspace methods have got a lot of research and application owingto their appealing properties, such as low time-consuming, well performance ondescription and separation. In this paper, we select the principal component analysis andthe local preserving projection algorithm based on subspace to analyze, research andimplement ate. To solve the problems found during the experiment, we improved theclassical algorithm. The performance of these algorithms has been improved to a certaindegree.The main work and contributions of the paper are summarized as follows:①The face images used in this paper from the existing face database, the facedetection process can be omitted. The pretreatment process is that images of the facedatabase used in crop, remove the excess part to retain the essential difference of theimage. Before the experiment we reduce the image pixel dimensions, reducing thecomputational complexity and removing the confounding factors of the image.②In the process of face recognition, the one-dimensional algorithm representedby vector with image matrix in turn and the two-dimensional algorithm computed byimage matrix directly are formed because of the different adoption of face images. Inthe second chapter one-dimensional principal component analysis method andtwo-dimensional principal component analysis method to analyze the instruction, verifythe strengths and weaknesses of the algorithm of two-dimensional algorithm byexperiment. One–dimensional algorithm is that the two-dimensional face matrix istransformed into one-dimensional vector operator, ignoring the internal structuralcharacteristics of the image, there by reducing the recognition rate. It calculates bydirect application of two-dimensional face matrix to retain the internal structure of theimage, but the higher the computational complexity of two-dimensional matrix.③Locality preserving projection algorithm is a typical method based on manifold learning. In the recognition process will encounter singular value problem affecting therecognition effect. In the third chapter a novel solution scheme by singular valuedecomposition (SVD) is proposed for recognition application. In the model, the sampledata is projected to a non-singular orthogonal matrix to solve the problem of singularvalue. Then obtain the data of the low dimensional sample space projection subspace,according to the locality preserving projection method. Project the training samples andthe testing samples on the low-dimensional subspace respectively. Finally use thenearest neighbor classifier for classification identification. In the experiment we verifythe validity and robustness of the algorithm.④An improved two-dimensional locality projection is proposed for facerecognition. Therefore, the nearest neighbor graph is constructed from all sample points,which each node corresponds to a column inside the matrix instead of the whole image,to better model the intrinsic manifold structure. In addition, the proposed improved2DLPP implements face images reduction to reduce the calculation complexity and finalfeature dimension.The experiments using the ORL face database and YALE face database to use thetext feature extraction algorithm and with the nearest neighbor classification method tobuild a complete face recognition program. We verify the effectiveness and stability ofthe algorithm using experimental comparison.