Cost-Sensitive Learning Based On Sparse Representation For Face Recognition

As one of the most popular research topics, sparse representation technique has beensuccessfully employed to solve face recognition task. Though current sparse representation basedmethods prove to achieve high classification accuracy, they implicitly assume that the losses of allmisclassifications are the same. However, in many real-world applications, differentmisclassifications could lead to different losses. Driven by this concern, we propose in this papertwo sparse representation based cost-sensitive algorithms for face recognition.We first propose a sparse representation-based cost-sensitive classifier (SRCSC). SRCSC usesprobabilistic model of sparse representation to estimate the posterior probabilities, separatelycalculates all the misclassification losses via the posterior probabilities and then predicts the classlabel by minimizing the losses. Our main contribution is to extend the sparse representationtechnique to cost-sensitive classifier.We then propose a cost-sensitive sparsity preserving projections (CSSPP). CSSPP considersthe cost information of sparse representation while calculating the sparse structure of the training set.Then, CSSPP employs the sparsity preserving projections method to achieve the projectiontransform and keeps the sparse structure in the low-dimensional space. In addition, we analyze therelationship between cost and the sparse coefficients.Experimental results on the public AR and FRGC face databases show that both of theproposed approaches can achieve high recognition rate and low misclassification loss, whichvalidate the efficacy of the proposed approaches.