Sparse Representation Robustness Algorithm For Occlusion Face Recognition

Recently sparse coding based on regression analysis has been widely used in face recognition research.Many researchers have noticed the relationship between different fidelity terms and the distribution of coding residuals.Most existing regression methods essentially simply use fidelity terms with l₁ norm or l₂ norm to indicate that the coding residual is a Gaussian distribution or Laplacian distribution,but in complex occlusion changes,the face image is not a single follow a certain distribution.In this thesis,two sparse representation robust coding with weight learning are proposed in combination with sparse representation robust coding,so as to effectively realize face image classification in multiple occlusion environments.The main contents are as follows:1.The research of the l₁ regularization minimization problem for sparse representation robust coding.Since the l₁ norm is not differentiable at the origin,the derivative of the l₁ norm cannot be directly obtained.In this thesis,the alternating direction method of multipliers is used to solve the l₁ regularization minimization problem.2.The research of the robustness of occlusion face recognition.First,the robustness of the algorithm is improved by improving the regression analysis model.On the one hand,the Huber function is used to make the fidelity term automatically match the l₁ norm or the l₂ norm;on the other hand,the l₁ norm regularization term is added to the coding coefficient,so that the coding coefficient is sparse.Secondly,by learning the sigmoid weight and the adaptive weight of coding residuals,the strong occlusion point can obtain a small weight value,so as to suppress the influence of occlusion on regression analysis and improve the robustness of the algorithm in occlusion face recognition.3.The research of the relationship between intra-class difference and the inter-class difference of face images.First,regression is performed by using a single category of sample subsets and query samples to reduce intra-class difference and avoid inter-class interference.Then the power exponent of sigmoid weight was changed to increase the relative difference between the intra-class difference and the inter-class difference.Finally,the adaptive weight are used to increase the negative correlation between the weight coefficient and the coded residual,thereby increasing the relative difference between the intra-class difference and the inter-class difference.At the same time,in adaptive weights,hyperparameters are less numerous and more interpretative.4.Based on the research of weighted sparse representation robust coding,this thesis elaborates the basic framework and main functional modules of the occlusion face recognition system,and finally designs a set of face recognition prototype system.