Research Of Face Recognition Algorithms Based On Subspace

The research for face recognition algorithms has been going on for decades, but at anearly stage the application of face recognition was limited as the low level of knowledge andcomputer technology. However, in the past two to three decades, with the rise of computerperformance and the need of social security, there has been many researchers who participatethe study of face recognition and make brilliant achievements at home and abroad. There aremany kinds of face recognition algorithms, among which the most widely used algorithm issubspace algorithm. Subspace algorithm classifies the training samples by statistical ideas,and mines useful local and global information which can be used to realize the mapping of thesubspace by solving optimal mapping function based on the Euclidean distance. In thesubspace, the samples can be classified easier.In this paper, we discuss two classical subspace face recognition algorithms, principalcomponent analysis and linear discriminant analysis. Principal component analysis is the firstsubspace algorithm applied to face recognition which removes the correlation betweensamples and preserves the energy of the original samples maximally. However, principalcomponent analysis does not contain the information of classification, so it exists manyrestrictions in the actual application. Linear discriminant analysis as a classical subspacealgorithm is very conducive to classification, it uses the information between samples andcenters of samples to classify the local and global information, and then realizes the optimalmapping in the subspace by mapping function. In the paper, we discuss how the selections ofenergy, number of training samples and dimensions of feature affect the classification.In this paper, we do lots of research on marginal fisher analysis method for theclassification between within-class information and between-class information. Marginalfisher analysis method classifies the information of within-class nearest samples andbetween-class nearest samples and realize that maximizing the distance of between-classsamples and minimizing the distance of within-class samples in the subspace through theoptimal mapping function. But marginal fisher analysis method requires two parameterswhich are difficult to adjust and have a great impact on the algorithm classification so that thealgorithm is hard to be applied in the real application. In the paper, we discuss how to selectthe number of between-class nearest samples to make the algorithm more robust, and proposea new method named spherical marginal fisher analysis method in which we selectbetween-class nearest samples based on the selection of number of within-class nearestsamples and don’t need other parameters to make the algorithm simpler and easier. Marginal fisher analysis method uses only the information between nearest samples, and don’t considerthe key role of the global information. In this paper, we propose a new algorithm based onglobal and local information, in which we prove the global information also plays veryimportant role in the face recognition algorithms and can combine with local information verywell. The experiments show that the two improved algorithms can effectively improve therecognition rate and are more robust than the original method.The main purpose of our study is to apply the algorithms to meet the needs of today’ssociety, and not only stays in the theoretical study, but it requires many considerations. In thispaper, we do a lot of experiment on all above methods, and analyze their performances onsubspace clustering, classification and time complexity so as to understand how to apply themin real application. Firstly, we use the degree of clustering of samples in the subspace to judgethe pros and cons of the choice of subspace, in which we verify that whether the within-classsamples cluster and the between-class samples go away or not. Secondly, we use the finalclassification results on the face databases to judge the classification effect of our methods.Finally, we calculate the performance time of all methods in the article to compare ourproposed method with the other methods.