The Research Of Feature Extraction In Face Recognition

Nowadays, face recognition is a hot issue in many research fields such as pattern recognition and computer vision, and it has been widely applied to the field of public security, authentication systems and video conference etc. Recognition algorithms first extract features from the training samples to get the feature space, then match the features of test samples and training samples through the classifier in test phase, finally output the recognition results. Therefore, the classifier design and feature extraction is the most critical issues in face recognition. The essence of feature extraction is finding a low-dimensional feature space which is more beneficial to classification, In the past few decades, many scholars have put forth many algorithms. However, these methods still have many deficiencies, and face image is very susceptible to outside interference, how to effectively improve the recognition rate and make full use of the image information is the main problem. In this paper, we do an in-depth research on some traditional methds based on subspace, and give some improvements, mainly related to the concepts of measure metric and l2norm. The experimental results on some face databases demonstrate the effectiveness of our presented methods.The main work for this thesis can be summarized as follows:1、Adaptive Supervised Discriminant ProjectionConsidering the weakness of Unsupervised Discriminant Projection (UDP), a novel feature extraction algorithm, termed Adaptive Supervised Discriminant Projection Analysis (ASDP), is proposed in the paper. The proposed algorithm does not only integrate a kind of dynamic feedback mechanism into construction of scatter matrices but also make full use of the class label information. The experimental results on AR and FERET face databases show the proposed method is more effective than some existing feature extraction methods such as UDP, LDA, etc.2、Similarity Preserving ProjectionWe know that each face image can be represented in terms of a linear combination of the eigenface by Principal Components Analysis (PCA), and PCA algorithm gives the best representation of images under the sense of minimum mean square error. However PCA only compares the Euclidean distance between projection coefficients of samples and ignores the residue between the original sample and its reconstructed one. Therefore a new concept called similarity distance metric is proposed in this paper. We project the two images into the same subspace and then characterize the similarity between pairs of samples by comparing both the projecting coefficients and the approximation errors simultaneously. The higher the value the more dissimilar the two samples, otherwise the more similar the two samples.Different from Locality Preserving Projections, A new method called Similarity Preserving Projections utilizes the concept of the similarity above and constructs the similarity scatter matrix, this algorithm doesn’t have to pre-set the number of neighbors, finally it gets the optimal projection subspace by maximizing the Objective function. The experimental results on AR and FERET face image database demonstrate the effectiveness of the proposed method.3、A Matching Pursuit Based Similarity Measure for Face RecognitionMatching pursuit algorithm as a sparse decomposition can not only uncover semantic information derives from a sample but important property of the data, but also have some advantages, such as simple, flexible and so on, this article takes it to select neighbors. All the training samples are used to build the overcomplete dictionary, we want to find the most relevant samples serve as a close neighbor of the sample. Next a new concept called similarity measure is proposed. We determine the weight of the neighbor matrix by comparing three factors:the ordered list of dictionary elements, the set of coefficients and the residue produced from matching pursuit approximation, finally it gets the optimal projection subspace by minimizing the Objective function. Compared with the other feature extraction method, the proposed method have a better recognition impact and more robust. The experimental results on AR and FERET face image database show that it is effective.4、Local Reconstruction Error of l2norm for Discriminant Feature ExtractionFor face recognition, learning robust and discriminative structure and feature extraction are the most important. Since Qinfeng Shi et al have demonstrated that l2approach to the face recognition problem is not only more accurate than the state-of-the-art method but also more robust, and much faster. Motivated by this, we propose a novel linear feature extraction method by combining the local discriminative power with l2norm technique. We calculate the local reconstruction weights using l2norm method. Finally, we seek a feature space, where the local reconstructive error caused by intra-class sample shall be as small as possible while the local reconstructive errors caused by data from different class but the most relevant shall be as large as possible. The properties make our method more effective and more robust than other methods, and the experimental results on AR and ORL face database demonstrate that the performance of our algorithm is the best.5、ocal Reconstruction Error and Measure Metric used in Face RecognitionFor two samples, let them project onto same projections. The relationship of the two samples is determined by the coefficients of the set and the reconstruction error at the same time. This is a new measure metric which construct the weight matrix of intraclass and the weight matrix of interclass. The representation coefficient is calculated by l2norm, then create the with-class local scatter of samples and the between-class local scatter of samples. Finally we seek to find a projection axis so that the samples from the same class are as close to each other as possible while the samples of different classes are from each other. This method can preserve the local similarity and the overall geometric structure simultaneously. The experimental results on AR, ORL and Yale face image database show that it is effective.