Research On Linear Representation Models For Face Recognition

Face recognition involves many disciplines such as image processing,pattern recognition,machine learning,and cognitive science.With the rise of artificial intelligence as a national development strategy,the value of face recognition reaserch is increasingly important.At present,the face recognition technology is flourishing,but in many application cases the inter-class correlation of face images is even much lower than the intra-class correlation,which causes the face recognition more complicated than other image classification tasks.In resent years,linear representation-based face recognition has achieved remarkable performance.These metheds assume that face images of the same category are lined in the same linear subspace and an unlabeled test sample is represented by the entire training set.We can add various constraints to get much better solutions to make the representation coefficients more conducive to classification.In fact,these representation-based algorithms are closely related to linear regression models in statistics.Therefore,the researches of linear representation-based face recognition focus on the development of regression models.In this paper,some specific researches and innovations are also performed on the loss function and regular terms inregression models,involving following works:1.We propose an Adaptive Class Preserving Representation?ACPR?based linear regression model.ACPR employs an improved nuclear norm to constrain the representation coefficients of samples in each class.This newly introduced regular term can adaptively select different constraints based on the structural information among the intra-class samples.When the correlation of samples in a class is high,ACPR tends to use a l₂-norm;when the correlation is low,ACPR degenerates to a l₁-norm based sparsity constraint.2.Adaptive Representation?AR?based regression model is proposed in this paper.AR model adds the overall regular term to constrain all training samples.Therefore,in the regression process,the AR model can not only use the intra-class structure information well,but also its overall constraint can adaptively take into account the global structure information of the training sample.3.In this paper,a robust loss function is introduced to improve the ACPR and AR models,and the above models are optimized by using the alternating direction method of multiplier?ADMM?.The face images in a real world application inevitably contains various changes and noises,so the l₁-norm loss function is more robust to the noise than the ordinary least squares method using its l₂-norm counterpart.