Improved two-step linear discriminant analysis method

Abstract

The invention relates to an improved two-step linear discriminant analysis method. The method specifically includes the following steps that: an approximate matrix method is adopted to solve the problem of the relatively high operation complexity of a TSLDA algorithm, an approximate eigenvalue matrix is adopted to replace an original eigenvalue matrix, so that the singularities of Sw and Sb are eliminated; and a screening and compression method is adopted to extract feature vectors in a TSLDA projection space which are favorable to classification, and single-feature Fisher information quantityis introduced to select feature vectors that meet conditions, and the selected feature vectors form an optimal projection space, thus, the projection space can be compressed, and the validity of allfeature information in the projection space can be ensured. The improved two-step linear discriminant analysis method has a high application value in fields such as image recognition including face recognition, fingerprint recognition, and handwriting font recognition.

Description

technical field [0001] The invention relates to the field of machine learning and image recognition, especially for the small sample problem in image recognition, and proposes a fast and high classification performance solution. Background technique [0002] Linear discriminant analysis (Lincar discrimination analysis, LDA) is a classic dimension reduction and feature extraction algorithm, widely used in face recognition, fingerprint recognition and gait recognition and other fields. The core idea of ​​LDA is to find an optimal projection space (transition matrix) W∈R for a set of linearly separable data d×h (h<d) After the sample is projected from the d dimension to the h dimension, the intra-class dispersion of the sample is the smallest, and the inter-class dispersion is the largest, that is, the samples in the projected space have the greatest separability. Its form is expressed as: [0003] [0004] where S w ∈R d×d is the intra-class dispersion matrix before ...

AI Technical Summary

Problems solved by technology

However, this method uses the cross-validation method to determine the regularization coefficient when eliminating the singular matrix, which increases the algorithm complexity and training time; at the same time, the feature information contained in the above four subspaces may not be completely effective, and may not contribute to the classification performance. Even the feature vectors (or noise information) that are not conducive to classification, how to filter out the feature vectors that contribute more to the classification performance is a problem that needs to be solved

Images