Research On Face Recognition Method Based On Global And Local Feature Fusion

Face recognition technology makes a relatively good progress and has initial applications in secure access control, video surveillance, content-based retrieval and next-generation machine interface areas and so on. But the current algorithm can not meet the actual demands of variety of applications in speed, accuracy, robustness. In recent years, the method based on local modeling has made great success. At the same time the modeling methodology based on the overall situation is developing rapidly. But the two methods in terms of speed, accuracy, robustness have their own advantages and disadvantages. In order to make good of use the effective aspect of these two methods, the face recognition method of fusion global and local characteristics is introduced by some scholars. As a novel face recognition method has inspired great interest in the academic circle and the field of application in face recognition.The present study will mainly focus on the extraction of global and local identification feature and the fusion of global and local identification, the main work of this paper is as following.(1) Considering the face detection, proposes a Face Detection algorithm based on a combination of SVM and AdaBoost. Then detect the face area in the sample and split out face based on the algorithm of detecting, which was used by feature extraction of face recognition.(2) Considering the extraction of global identification feature, extract the overall identification characteristics of face region use the method based on Kernel principal component analysis (KPCA).First of all, extract global identification feature of the face region split by the above method. This method avoids inconvenience in the calculation of inner product in high dimensional high feature, and does not need to solve nonlinear optimization problem, compared with the other methods , can improve feature extraction efficiency and speed for only involves decomposition matrix eigenvalue calculation.(3) Considering the extract of local identification, extract local feature of face region use the method of block kernel impendency principal component analysis (BKICA).The method will be divided into uniform small block as a local area firstly, extract identifying characteristics of the local area secondly; then proposes the strategy of fuse this characteristic with adaptive weighted algorithm according to the degree of personalization of local features. This method effectively reduces the small sample of the problem of adverse effect, but also reduces the sample space of dimension and since the process of the complexity of more recognition rate.(4) Considering the fusion of global identification and local identification,proposes a novel face recognition method of fusion global and local characteristics of face based on improved D-S(IDS) .Considering factors of light, expression (eyes, shielding, scarf) and so on, a method of face recognition of fusion of global and local is proposed. While extracting global identification feature of the whole face image, at the same time extracting local area identification characteristics, conclude ultimate recognition result by the fusion global identification similarity and local identification similarity based on the principle of improved D-S evidence theory finally.Some experimental results on JDL and AR face databases show that the methods in this thesis have higher adaptability and recognition rate compared with the global feature identification method based on KPCA and the local feature identification method based on KICA.