Stacked Face De-Noising Auto Encoders for Expression-Robust Face Recognition

Abstract

Recent advancement in unsupervised and transfer learning methods of deep learning networks has seen a complete paradigm shift in machine learning. Inspired by the recent evolution of deep learning (DL) networks that demonstrates a proven pathway of addressing challenging dilemmas in various problem domains, we propose a novel DL framework for expression-robust feature acquisition. The framework exploits the contributions of different colour components in different local face regions by recovering the neutral expression from various expressions. Furthermore, the framework rigorously de-noises a face with dynamic expressions in a progressive way thus it is termed as stacked face de-noising auto-encoders (SFDAE). The high-level expression-robust representations that are learnt via this framework will not only yield better reconstruction of neutral expression faces but also boost the performance of the subsequent LDA[1] classifier. The experimental results reveal the superiority of the proposed method to the existing works in terms of its generalization ability and the high recognition accuracy.