Affiliations: [a] School of Mechanical and Electrical Engineering, Changchun University of Technology, Changchun, China
| [b] Jilin Communications Polytechnic, Changchun, China
| [c] School of Electronics and Electrical Engineering, Changchun University of Technology, Changchun, China
| [d] School of Artificial Intelligence, Jilin University, Changchun, China
| [e] School of Electrical and Information Engineering, Jilin Engineering Normal University, Changchun, China
Correspondence:
[*]
Corresponding author. Xiaomei Lin, School of Electronics and Electrical Engineering, Changchun University of Technology, Changchun 130012, China. E-mail: 187049860@qq.com.
Abstract: As a deep learning network model, ResNet50 can effectively recognize facial expressions to a certain extent, but there are still problems such as insufficient extraction of local effective feature information and a large number of parameters. In this paper, we take ResNet50 as the basic framework to optimize and improve this network. Firstly, by analyzing the influence mechanism of the attention mechanism module on the network feature information circulation, the optimal embedding position of CBAM (Convolutional Block Attention Module) and SE modules in the ResNet50 network is thus determined to effectively extract local key information, and then the number of model parameters is effectively reduced by embedding the depth separable module. To validate the performance of the improved ResNet50 model, the recognition accuracy reached 71.72% and 95.72% by ablation experiments using Fer2013 and CK+ datasets, respectively. We then used the trained model to classify the homemade dataset, and the recognition accuracy reached 92.86%. In addition, compared with the current more advanced methods, the improved ResNet50 network model proposed in this paper can maintain a balance between model complexity and recognition ability and can provide a technical reference for facial expression recognition research.
Keywords: ResNet50, SE, CBAM, depth separability, lightweight
