Face Classification Across Pose by Using Nonlinear Regression and Discriminatory Face Information

Face recognition across pose cripples with the issue of non-availability of few important facial features. Some of the facial key features undergo occlusion during pose variations. The sole application of linear regression model in face recognition across pose is unable to predict the occluded features from the remaining visible features. With the approach like discriminative elastic-net regularization (DENR), the training sample’s discriminatory information, is embedded into regularization term of the linear regression model. Classification is realized using least sqaure regression residuals. However, the existence of nonlinear mapping between frontal face and its counterpart pose limits the application of DENR. In this paper, discriminative elastic-net regularized nonlinear regression (DENRNLR) is proposed for face recognition across pose. DENRNLR learns discriminant analysis-based kernelized regression model constrained by elastic-net regularization. The effectiveness of the proposed approach is demonstrated on UMIST and AT&T face database.