Isometric mapping based facial image recognition method

Abstract

An isometric mapping based facial image recognition method belongs to the technical field of image processing and comprises the following steps: processing the input images and expressing the images in the form of vectors; taking column vectors as training samples to constitute a training sample set, and then taking each training sample in the training sample set as a vertex to construct a neighborhood linked undirected weighed graph GX; computing the shortest path between every two training samples according to the neighborhood linked undirected weighed graph GX and establishing the shortest path matrix DG=(dG (i, j)); carrying out direct linear classification processing on the shortest path matrix and obtaining the optimal projection matrix W; and acting the optimal projection matrix W on a test set Dtest to reduce dimensions, thus obtaining the low-dimensional optimized feature samples, and then classifying the test set by the minimum distance method or the K-nearest neighbor method to obtain the face recognition results. The method simultaneously takes the features of the spatial relations between image pixels into consideration, uses direct linear discriminant analysis to replace multi-dimension analysis and obtains higher recognition rates in the experiments of face recognition.

Description

technical field [0001] The present invention relates to a method in the technical field of image processing, in particular to a face image recognition method based on isometric mapping. Background technique [0002] In traditional pattern recognition, classical subspace methods such as Principal Component Analysis (PCA), Fisher Linear Discriminant Analysis (Fisher Linear Discriminant Analysis, LDA) and so on essentially use various optimization criteria to find high-dimensional input spaces to low-dimensional ones. The optimal projection matrix of dimensional subspace, they have achieved good results in applications such as face recognition and text recognition. But when encountering some complex nonlinear problems, they lose their advantages. Based on this, nonlinear methods emerge as the times require, one is a method based on kernels, and the other is a method based on manifold learning. Manifold learning methods based on spectral theory can be divided into local methods...

AI Technical Summary

Problems solved by technology

This point-based operation is highly sensitive to noise, especially in the field of image recognition, even a small deformation will affect the processing effect of the entire method, and it is not topologically stable.

Images