Dimensionality Reduction And Recognition Technology Of Digital Image-algorithm Research Of Face Recognition

In recent years,face recognition research become a highly valued and active research topic in the fields of computer vision and pattern recognition because of its huge research prospects and potential development in authentication, file management and visual communications. General face recognition system consists of five functional modules which contains information access, digital and pre-processing, feature extraction, classifier design and training, pattern classification. Feature extraction is the most critical and core part and one of the biggest difficulties in the five modules of the system. To achieve the transformation of raw data through the mapping from the face image space to feature space, and get the most essential features which reflects the classification, and provide the basis of next effective classification and recognition. The effectiveness and stability of feature extraction is directly related to the recognition effect.The author of this paper made some in-depth study in several feature extraction of face recognition, the main idea are as follows:First, the non-parametric subspace analysis which is based on linear discriminant analysis method is proposed in 2009, and it is a new face recognition algorithm. The biggest difference from linear discriminant analysis method is the definition of the between-class scatter matrix, the new between-class scatter matrix has consider the center of class , but also take the boundaries of class structure into account which has great significance of the classification. But the within-class scatter matrix Sw of the NSA is the same like the LDA, and not considers that selecting different KNN points will produce different between-class scatter matrix in calculating the between-class matrix. Author proposed a new method based on the NSA method. The original images are divided into modular sub-images then NSA is utilized on the new pattern which is obtained by modular 2DPCA to extract the final features from the sub-images. The experimental results obtained on the facial database ORL ,XM2VTS and AR, show that the recognition performance of the new method is superior to that of the primary method of LDA and NSA.Secondly, the kernel method is an effective method of dimension reduction, but also a usual statistical feature extraction method. Our ability to solve nonlinear problems because of the use of kernel method has greatly improved. KDA algorithm projects the high-dimensional of original data samples to the low-dimensional feature space through non-linear mapping, and then uses the LDA method to extract feature. It is applied in face recognition area, achieved high recognition rate. However, kernel methods still have flaws and shortcomings, use different kernel function in the same the database will get very different recognition rate, and a reasonable kernel function can achieve excellent recognition result than any other kernel function. The standard to determine a kernel method is selecting what kind of kernel function, and kernel function is the key step to solve the problems. A new method is proposed based on kernel discriminant analysis ,it added a weight function in the between-class scatter matrix, The experimental results obtained on the facial database ORL and XM2VTS unable to show higher recognition rate than the comparative method , However, the stability of the algorithm has a better performance.