基于稀疏近似的特征选择及人脸识别

Inspired by biological vision systems, the over-complete localfeatures with huge cardinality are increasingly used for facerecognition during the last decades. Accordingly, feature selectionhas become more and more important and plays a critical role forface data description and recognition. In the paper, atrainable feature selection algorithm based on the regularized framefor face recognition is proposed. By enforcing a sparsity penalty term on theminimum squared error (MSE) criterion, the feature selectionproblem is casted into a combinatorial emph{sparse approximation} problem,which can be solved by greedy methods or convex relaxation methods.Moreover, based on the same frame, a emph{sparse}Ho-Kashyap (HK) procedure is proposed to obtain simultaneously the optimalemph{sparse} solution and the corresponding emph{margin} vector ofthe MSE criterion. The proposed methods are used for selecting themost informative Gabor features of face images for recognition andthe experimental results on benchmark face databases demonstrate theeffectiveness of the proposed methods.