Feature Extraction Technology And Its Application To Face Recognition

Feature extraction or dimension reduction is the primary problem in pattern recognition. Its basic task is to seek the most effective features for classification from all the source features. This paper has a deep research of linear dimension reduction and nonlinear dimension reduction techniques and proposes some modified algorithms for face recognition. The main work is followed:1. A modified principal component analysis method is proposed. This method uses the within-class sample covariance matrix as the producing matrix, and makes use of average sample vector and between-class sample covariance matrix to select eigenvectors after PCA transformation. So the label information of the projected samples can be kept as much as possible.2. Based on the locally linear embedding algorithm and considering the mal-distribution of samples, a new distance is adopted to change their corresponding distance according to the area situation they belongs to. So the affect caused by k value is reduced. And a new solution for obtaining the low-dimensional feature vector for test sample is proposed.3. Combining the improved principal component analysis and locality preserving projection methods, a modified LPP method is generated. This method has both the supervised feature and manifold study advantages. The low-dimensional feature vector obtained has considered the non-linear manifold structure and discriminating ability...