Face Recognition Algorithm Based On Low Rank And Sparse Representation

The current face recognition methods often ignore the noises during training process.When training samples corrupted, recognition performance will significantly degenerate.However, recently proposed low rank decomposition can effectively ease thisphenomenon. On the basis of the relevant research results and the latest research progress,this paper studies the robust face recognition algorithm.Firstly, according to the station of low-rank decomposition can effectively suppressnoise sparse in training samples, this article explores four typical low rank decompositionalgorithms for face recognition and compares the performance of each algorithm.Secondly, this article studies a novel low rank metaface dictionary face recognitionalgorithm which combined with sparse representation. This method uses metafacedictionary to represent test data in transform space and classify the test sample by thelowest reconstructed error method. Experiment results show that the proposed algorithmhas a higher recognition rate.Then, since relaxed collaborative representation can get more discriminating codingcoefficients, this text explores a novel low rank relaxed collaborative representationalgorithm which combined with global and local features. This method uses blocking anddown sampling to get the local features and global features of the training and testingsamples, then, obtains training dictionaries in transform space by low-rank decompositionand principal component analysis. Next, this method carries out the process of relaxedcollaborative representation in transform space to classify test data according to the lowestweight coded residuals. Experiment results show that the proposed algorithm has a higherrecognition rate and better robustness.Finally, the existing face recognition algorithms can not guarantee the recognitionrate and speed at the same time. To solve the problem, this article studies a novel low rankblocked collaborative representation algorithm which combined with global and localfeatures. Similarly, this method uses blocking and down sampling to get the local featuresand global features of the training and testing samples. After get training dictionaries by low-rank decomposition and principal component analysis and test data in transform space,the proposed algorithm combines the voting method with the lowest reconstructed errormethod to classify. Experiment results show that the proposed algorithm has a higherrecognition rate and faster recognition speed.