Extreme learning machine face dimension reduction method based on discriminant shared neighborhood preservation

Abstract

The invention discloses an extreme learning machine face dimension reduction method based on discriminant shared neighborhood preservation. The method comprises the following steps: S1, image preprocessing: performing scale normalization processing on a face image acquired by a face image acquisition module; s2, an input layer: inputting a face feature X subjected to scale normalization processing; s3, a hidden layer: selecting a nonlinear excitation function, randomly and nonlinearly mapping the single-sample face feature X to an N-dimensional feature space, and transforming to obtain a feature H (X); s4, outputting a layer: fixing a feature H (X); and S5, a sample data storing the face sample data t subjected to dimension reduction processing. The invention provides an SGRD-ELM algorithm. The SGRD-ELM not only considers the category information of the sample, but also pays attention to mining the shared neighbor information of any two sample points, and under the condition that the dimension of the sample data is reduced to 64 dimensions, the human face recognition rate is averagely improved by 10%-14% compared with that of an LLE method, and the human face recognition rate is averagely improved by 6%-8% compared with that of a US-ELM method.

Description

technical field [0001] The invention belongs to the field of biometric identification systems, in particular to an extreme learning machine face dimension reduction method based on discriminative shared neighborhood preservation. Background technique [0002] With the advent of the era of artificial intelligence and big data, people need to verify identity information in daily life, economic trade, information registration and inspection. Nowadays, identity authentication technology has been widely used in various fields of the country and society. In recent years, with the promotion of electronic banking, online commerce, online education, intelligent access control and other models, people are more and more concerned about the security of identity information. Pay attention to. Traditional identity authentication methods (such as passwords, ID cards, etc.) have been unable to meet people's needs for personal information security because they are easy to steal and counterf...

AI Technical Summary

Problems solved by technology

[0006] 1. Linear dimensionality reduction technology cannot mine and analyze the potential geometric structure of the data. For nonlinear high-dimensional data with low-dimensional manifolds, the dimensionality reduction performance will be greatly reduced

[0007] 2. Non-linear dimensionality reduction methods such as kernelized linear dimensionality reduction and manifold learning do not have a clear explicit mapping expression, and cannot solve out-of-sample problems, that is, when new samples are added, these nonlinear dimensionality reduction models need to be relearned and optimized

[0008] 3. The US-ELM algorithm has poor ability to mine high-dimensional nonlinear data sample point neighborhood information and cannot consider category information

Images