Lightweight End to end Pose-Robust face recognition system with Deep Residual Equivariant Mapping

Abstract

Abstract In the face recognition field of study, pose-robustness and lightness of a model are few of the critical improvement factors of face recognition. However, these fields are still providing challenge for researchers. Even though pose variance is proven to drop the accuracy of deep learning-based models, pose-robustness is not studied often in lightweight face recognition models. Existing pose-robust models have heavier implementation costs compared to lightweight models. We propose a deep learning architecture that implements Deep Residual Equivariant Mapping (DREAM) to improve pose-robustness of a lightweight MobileFaceNets model as a solution to the underlying issue. In the proposed model, the DREAM block is stitched to the MobileFaceNets stem CNN architecture. The evaluation process compares the speed, file size, and accuracy on pose diverse datasets, such as the CFP and IJB-A dataset. The evaluation results of the proposed model show an accuracy improvement of 0.07% with verification speed difference of 0.17 ms. Both of the results show a better performance compared to the baseline naive model.