Based On Kernel Feature Fusion And Selection Of Face Recognition Research

As a research subject possessing high theoretical and practical value, face recognition has been an important branch of pattern recognition and computer vision. Feature extraction is the key point in face recognition, and different extraction can result in different features. So currently comprehensive application of different forms of feature is a hot research spot in face recognition. One important way of solving the problem is feature fusion, which can identify and improve the recognition efficiency and accuracy rate by integrating the different features. In addition, since the extracted facial features are often high-dimensional vector, it not only makes the algorithm computational complexity (including the computing time and storage space) large, but also the possible redundancy in the features may reduce the algorithm's recognition performance. Therefore the selection of face recognition becomes necessary. Thus it's extremely important to study the approaches of feature selection.This paper focuses on the face recognition feature fusion and feature selection problems, mainly from the following aspects of work:(1)Face image pre-processing is the prerequisite for feature extraction and recognition. Precise positioning of the human eye is an important part of pre-processing stage, Images are cropped according to the human eye to achieve geometric normalization, and then complete image pre-processing through gray normalization. During the course of human eye orientation, projection positioning will be the focus of analysis. A hybrid projection peak analysis of the human eye location is proposed. By mixing variance projection function with the gradient projection function as well as analysis of projected peak in the projection curve, the exact location of the human eye are determined.(2)Linear subspace and kernel methods of human face feature extraction are described in detail together with the advantages and disadvantages.Experimental research concentrates on the kernel function of kernel methods and parameter settings.(3)For the shortage in feature extraction of unsupervised KPCA and supervised null space KFDA, a feature fusion of combining the two features into account is proposed, which is the integration of serial feature fusion and parallel feature fusion, and the use of it to classification and recognition. According to the ORL and Yale face database, experimental results show that with the help of third-order neighbor classifier, feature fusion method avoids the shortcomings of both recognition performance and is superior to other algorithms.(4)By the relationship between KPCA algorithm error rate and feature dimension changes, feature redundancy problems appear when using KPCA feature extraction algorithm for recognition, hence there is the need for feature selection. In view of the advantages in dealing with combinatorial optimization problems, through well-designed fitness function, discrete binary particle swarm algorithm can be used for KPCA feature space for feature selection, and search for the optimal feature subset. Experimental results on the ORL and Yale face database show that by the use of third-order neighbor classifier, the integration of discrete particle swarm optimization in KPCA algorithm has higher recognition accuracy, which is superior to the KPCA method.