Cross-Domain Facial Expression Recognition Based on Transductive Deep Transfer Learning.

Keyu Yan,Wenming Zheng,Tong Zhang,Yuan Zong,Chuangao Tang,Cheng Lu,Zhen Cui

DOI: https://doi.org/10.1109/access.2019.2930359

IF: 3.9

2019-01-01

IEEE Access

Abstract:In this paper, we proposed a novel end-to-end transductive deep transfer learning network (TDTLN) to deal with the challenging cross-domain expression recognition problem, in which both the source and target databases are utilized to jointly learn optimal nonlinear discriminative features so as to improve the label prediction performance of the target data samples. As part of the network parameters, the labels of the target samples are also optimized when optimizing the parameters of TDTLN, such that the cross-entropy loss of source domain data and the regression loss of target domain data can be simultaneously calculated. Finally, to evaluate the recognition performance of the proposed TDTLN method, we conduct extensive cross-database experiments on four commonly used multi-view facial expression databases, namely the BU-3DEF, Multi-PIE, SFEW, and RAF database. The experimental results show that the proposed TDTLN method outperforms state-of-the-art methods.