Multi-pose face reconstruction and Gabor-based dictionary learning for face recognition

Many methods based on data and model analysis have been proposed to overcome the adverse effect of poses, occlusion, and illumination on face recognition. Frontal face generation from multi-pose faces is still a widely studied and challenging problem due to its ill-posed. We focus on using a simple model to overcome the effect of pose changes on face recognition. This paper proposes a multi-pose face reconstruction model (MPFR) to generate available face information and combines the model with Gabor-based dictionary learning methods to learn discriminant features. The MPFR is adopted to generate the frontal face image. Given the distortion of the image only generated by Generative Adversarial Networks, the identity loss functions and the symmetry loss function are utilized in the multi-pose faces reconstruction model to reconstruct a more realistic reconstruction frontal image. Besides, we combine discriminative dictionary learning with Gabor features to better express face features for image classification. We report qualitative visualization results and quantitative recognition results of the MPFR model. Further, the experimental results demonstrate the application of the MPFR model in face recognition.

Automated summary

This paper proposes a multi-pose face reconstruction model (MPFR) to generate available face information and overcome the effect of pose changes on face recognition. The MPFR model combines Gabor-based dictionary learning methods to learn discriminant features. The experimental results demonstrate the application of the MPFR model in face recognition.