Research On Face Recognition Algorithm Based On Sparse Representation

With the development of science and technology,people have paid more and more attention to face recognition and related technologies.Its theoretical research promotes the coordinated development of many fields,and face technology develops rapidly.Although face recognition technology is becoming more and more mature,there are still many problems.Such as illumination conditions,sample size,sample attitude and other factors will affect the performance of the recognition algorithm.Researchers have designed a variety of algorithms to solve various problems,among which the algorithms based on sparse representation and low-rank representation have particularly attracted the attention of researchers.In this paper,we design two improved algorithm models based on low rank representation and sparse representation respectively for small samples and sample margins.(1)The traditional zero-one matrix is used as the regression target for most linear regression methods,which greatly reduces the flexibility of the regression model.In this paper,an elastic net regularized linear regression model is designed,which is called the elastic net regularized linear regression model with discriminant,which combines the elastic net regularized linear regression with negative ? drag technique.The model is compared with other linear regression models.First,we extend the boundaries of the different strategies by relaxing the strictly binary objective to a more feasible matrix of variables.Secondly,we use robust singular values of the elastic net to enhance the compactness and validity of the projection matrix.At the same time,the negative ? drag technique is used to classify the noise and pollution data robustly.The basic principle of the negative ? drag technique is that the expected generalization ability can be obtained by the relatively small classification margin of the least square regression training process.Compared with traditional least square regression,the class boundary of negative drag technique is reduced to obtain robust results when classifying noise data.The whole model has a strong discriminability,which provides a more robust classifier for the deformed and noise-polluted data,thus obtaining a better classification effect.Therefore,it is very helpful to improve the classification performance.(2)For the high dimension of face data and the insufficient number of samples.In this paper,we propose a feature extraction algorithm which combines increasing virtual sample and collaborative representation,and it is called virtual sample enhanced collaborative projection algorithm.First,we take advantage of the "axial symmetry" of the face to generate two virtual face images from the original face in an iterative way,and integrate the original face and virtual face into the required sample data set.Finally,we aim to maintain the cooperative reconstruction relationship of these sample data by minimum-norm regularization correlative objective functions.At the same time,considering the existing flow structure in face data,we use ridge regression to find the optimal weight matrix.This algorithm weakens the influence of small sample data set and nonlinear factors in face recognition process.Finally,the experiments on each face database show that the proposed algorithm has satisfactory performance compared with other traditional unsupervised dimension reduction algorithms.