Face Recognition Algorithms Research Based On Manifold Learning

At present, face recognition is one of the most important and active research topics in the pattern recognition and artificial intelligence fields. Compare with other biological features, for example, fingerprint, iris and voice, face has some advantages such as direct, security, impossible to copy, convenience and non-contact. Because of this, face recognition technology has become one of the fastest develop and the most potential techniques in many biological recognition technologies and it is widely used in security check criminal investigation, information security and other fields. However, face recognition is still facing many difficulties and challenges. Firstly, face images are susceptible to the external factors such as illumination, expression, posture and imaging angle. Secondly, high dimensional face data restricts the speed and accuracy of face recognition. The research of this paper is based on manifold learning, which is around the feature extraction, feature dimension reduction of face recognition, and the main work includes:1. Facial expression analysis based on wavelet decomposition and manifold learning. As everyone knows, face samples dispersed in the high-dimensional space usually produce complex computing problems. In order to solve this problem, the concept of image granularity is applied at the feature level. By this strategy, the dimensionality of high-dimensional data is reduced and computational complexity of the algorithm decreases efficiently. The distribution and changes of the high-dimensional face pose and facial expression are presented. In additional, the runtime and the energy of low frequency subband after different levels of wavelet decomposition are analyzed.2. Research of feature dimension reduction based on Gabor wavelet. To solve the problem that the face recognition is easily affected by external factors such as illumination, expression, this part mainly introduces the two improved methods proposed in this paper. One is supervised 2DNPE method based on Gabor wavelet in face recognition; another is supervised B2 DLPP approach based on Gabor wavelet in face recognition. These two algorithms are both supervised extension of the original methods. By making full use of the class information, the discriminating ability of these two algorithms are further improved. In addition, the images obtained by Gabor filter usually are able to overcome the impact of external factors such as illumination, scale, expression and so on. Therefore, Gabor wavelet is used to extract features from face images in the phase of feature extraction to make the proposed approaches in this paper be more robust to those changing factors. By comparing the classification performance between the proposed methods and other classical face recognition algorithms, the effectiveness and superior of the proposed methods are demonstrated by the experimental results.3. Research of bilateral two-dimensional neighborhood preserving discriminant embedding method. The bilateral two-dimensional neighborhood preserving embedding algorithm is a kind of classical two-dimensional manifold learning methods, which reduces the dimension of high-dimensional data from row and column directions by keeping the reconstruction weights of near point within the neighborhood be the same. However, B2 DNPE algorithm is an unsupervised method. Therefore, we proposed a novel method named bilateral two-dimensional neighborhood preserving discriminant embedding(B2DNPDE) to improve the discriminating ability of B2 DNPE. In B2 DNPDE, by fully considering the within-neighboring and between-neighboring information, the optimal projection matrix can be computed by minimizing the intra-class scatter and maximizing inter-class scatter based on Fisher criterion. The performance of the proposed method is evaluated on the Yale, PICS and AR three public face databases and the experiment results show the validity of B2 DNPDE.