**5. Conclusions**

This paper proposed a facial expression recognition model based on deep transferred learning and a novel weighted-cluster loss to mitigate the shortage an imbalanced data problems. An SE-ResNet-50 model which is pre-trained for face identification task is fine-tuned to recognize eight common facial expressions in AffectNet data. This not only helps to save computing resource but also alleviate the shortage of training data problem. Then, the proposed weighted-cluster loss was used in the fine-tuning phase to tackle the high imbalance in data distribution of AffectNet data. Multiple metrics have been used to evaluate the effectiveness of the proposed model. Experimental results on the test set indicate that the proposed FER model can outperform its counterpart models which uses either weighted-softmax loss or center loss. However, the proposed model is built to recognize facial expressions on static image data, which may limit its applicability. Moreover, using generative adversarial networks (GAN) to generate more data for training FER models is considered as our other future work.

**Author Contributions:** Conceptualization, Q.T.N. and S.Y.; Data curation, Q.T.N. and S.Y.; Formal analysis, Q.T.N. and S.Y.; Funding acquisition, S.Y.; Investigation, S.Y.; Methodology, Q.T.N. and S.Y.; Project administration, S.Y.; Resources, S.Y.; Software, Q.T.N.; Supervision, S.Y.; Validation, Q.T.N. and S.Y.; Visualization, Q.T.N. and S.Y.; Writing—original draft, Q.T.N. and S.Y.; Writing—review and editing, Q.T.N. and S.Y. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by the 2020 Research Fund of University of Ulsan.

**Conflicts of Interest:** The authors declare no conflict of interest.
