Research On Face Recognition Algorithms Based On Sparseness And Low-rank Constraints

Representation learning becomes a chronically active research direction in the field of machine learning and computer vision in recent years,and plays an indispensible role in these fields.Especially,the well-defined representation is critical to achieving satisfactory performance of learning systems.Sparse representation(SR)and low rank(LRR)have become very popular and widely concerned of representation learning.So far,many researchers have proposed a large number of improved algorithms based on low rank and sparse representation,and those algorithms have obtained satisfactory results in practical applications.In this paper,we discuss some classical and recent algorithms based on low rank and sparse constraints,and analyze the shortcomings of these algorithms.This dissertation proposes two novel improved algorithms for image recognition.(1)We propose a discriminative elastic-net regularized representation learning(DENRL)method for robust image recognition.We first introduce a robust subspace learning model with the elastic-net regularization of single values to obtain a compactness data representation.Then,the proposed method builds a bridge between the training samples and the test samples in the framework of semi-supervised learning,which ensures that the representation is consistent between them.Moreover,we impose constraints on the subspace model which can enhance the connections within the same class and eliminate the correlations between different classes.Finally,we optimize the algorithm by the alternating direction method to improve the computation efficiency.The extensive experiments show that the proposed method can obtain higher recognition rates in comparison with the state-of-the-art recognition methods.(2)Sparse representation still faces with many challenging problems.For example,the existing sparse representation methods cannot make full use of the internal information of the global structure of the data.In addition,the sparse representation algorithm cannot get satisfactory data representation,when the dimension of the feature is higher than the number of samples.Therefore,we propose a simple and robust dimensionality reduction method based on CRC framework.First,we use pooling to reduce the dimension of the features,which ensures that we can get the most useful information and eliminate noise.Then,we use the adaptive weighted fusion method to fuse the residuals which obtained from different sources,so that a robust image classification result is obtained.Finally,extensive experimental results show that the proposed method is robust,simple and efficient.In this paper,two improved algorithms based on low rank and sparse representation are proposed.We selected several common public face datasets to verify the superior performance of the proposed algorithms.The databases contain AR face database,Yale B face database,ORL face database and Fifteen Scene Categories database,etc.Furthermore,we compare with state-of-the-art recognition algorithms on each dataset.Extensive experiments show that the proposed methods have a higher recognition rate than the existing recognition methods.