Kernel Sparse Description Face Recognition Method Based on Geodesic Mapping Analysis

Abstract

The invention discloses a kernel sparse description face recognition method based on geodesic mapping analysis. The unique geodesic line connected to the Riemannian manifold is obtained through two pixel points, and the geodesic line is projected to the tangential direction through logarithmic and exponential mapping. Space, so as to obtain more discriminative classification features, can better adapt to the special surface structure of the face, can accurately measure the real distance between two pixels of the face image under the condition of expression, posture and heavy occlusion, through The nonlinear mapping converts the data vector in the tangential space into a higher-dimensional feature space, introduces the kernel function to model the sparse feature space, and solves the optimization problem of the L1 norm in the nonlinear sparse space to achieve accurate classification of faces , which greatly reduces the impact of complex condition changes on recognition performance, makes the algorithm more robust, and solves the problem that the existing face recognition system is susceptible to changes in posture, expression and occlusion in an uncontrolled environment.

Description

Technical field: [0001] The present invention relates to the field of face recognition, in particular to a method for face recognition based on sparse kernel description of geodesic mapping analysis. Background technique: [0002] Face recognition is a research hotspot in the field of machine vision and artificial intelligence, and it is also an important authentication method for information security and social security. In recent years, a large number of research literatures project high-dimensional test face images into low-dimensional feature spaces, such as Eigenfaces, Fisherfaces, Laplacianfaces, etc. All these algorithms require test samples Must be properly cropped, aligned, and of the same scale as the training samples. However, the actual outputs of face detectors are neither aligned nor constrained, and vary widely in scale, making it very challenging to classify and identify these data. [0003] Sparse description has received great attention in the field of ta...

AI Technical Summary

Problems solved by technology

[0004] In addition, most of the literature uses Euclidean distance as the face similarity measure, but it has a large error when the face pose and expression change

Images