Research Article
A Bi-directional Residual Network for Image Expression Recognition
@INPROCEEDINGS{10.1007/978-3-030-72792-5_16, author={Daihong Jiang and Sanyou Zhang and Cheng Yu and Chuangeng Tian}, title={A Bi-directional Residual Network for Image Expression Recognition}, proceedings={Simulation Tools and Techniques. 12th EAI International Conference, SIMUtools 2020, Guiyang, China, August 28-29, 2020, Proceedings, Part I}, proceedings_a={SIMUTOOLS}, year={2021}, month={4}, keywords={Residual network Deep learning Image recognition Feature expression}, doi={10.1007/978-3-030-72792-5_16} }- Daihong Jiang
Sanyou Zhang
Cheng Yu
Chuangeng Tian
Year: 2021
A Bi-directional Residual Network for Image Expression Recognition
SIMUTOOLS
Springer
DOI: 10.1007/978-3-030-72792-5_16
Abstract
In this paper, an improved model based on the combination of residual and inverted residual blocks is proposed for image expression recognition, named as bi-directional residual network. The main objective of the proposed method is to alleviate the problem of feature dispersion due to the deep network level in traditional expression recognition research. In this case, residual block is a good solution. However, residual network with small scale of training data can easily lead to over-fitting, which is often the case for image expression recognition. To improve the robustness of the network during training, inverted residual blocks are therefore adopted. Depending on the organization sequence of residual blocks and inverted residual blocks, three network structures are proposed and studied. Fer2013 and CK+ datasets in facial field are adopted for experiment. The experimental results show that the optimized algorithm improves the accuracy by 2.79% on Fer2013 dataset compared with ResNet-50 models.
