下载全文： 202305013.pdf

文章编号： 1672-6987（2023）05-0109-06； DOI： 10.16351/j.1672-6987.2023.05.013

刘宁1， 孙萍2， 冯宇平1*， 鞠伯伦3(1.青岛科技大学 自动化与电子工程学院， 山东 青岛 266061； 2.青岛海湾化学股份有限公司，山东 青岛 266409；

3.中国船舶集团有限公司第七一六研究所，江苏 连云港 222006)

摘要： 为了避免CNN池化层丢失特征并解决胶囊网络自身卷积层特征提取不足的问题，提出一种改进的ResNet18和胶囊网络结合的人脸表情识别方法。该方法仅保留ResNet18的卷积层，对其中的3个残差块进行调整，然后融入CBAM注意力机制，替换胶囊网络的单卷积层来提取特征，最后将提取的特征送入胶囊网络进行训练和分类。所提出的方法在CK+、RAF-db和FER+数据集上进行了实验，识别准确率CK+9797%、RAF-db8411%、FER+（单标签）8624%、FER+（双标签）9414%。实验结果表明，该方法在人脸表情识别方面具有可行性和有效性。

关键词： 胶囊网络； ResNet18； 人脸表情识别； 特征提取； CBAM注意力机制

中图分类号： TP 391.4文献标志码： A

引用格式： 刘宁， 孙萍, 冯宇平, 等. 基于ResNet18与胶囊网络结合的人脸表情识别［J］. 青岛科技大学学报(自然科学版), 2023, 44(5): 109-114.

LIU Ning, SUN Ping， FENG Yuping， et al. Facial expression recognition based on the combination of ResNet18 and capsule network［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition), 2023， 44（5）： 109-114.

Facial Expression Recognition Based on the Combination of

ResNet18 and Capsule Network

LIU Ning1, SUN Ping2, FENG Yuping1, JU Bolun3

(1.College of Automation and Electronic Engineering, Qingdao University of Science and Technology， Qingdao 266061,China;

2.Qingdao Haiwan Chemical Co.,Ltd., Qingdao 266409, China; 3.The 716th Institute of CSSC, Lianyungang 222006, China)

Abstract: In order to avoid the problem of CNN pooling layer missing features and solve the problem of insufficient feature extraction of the convolutional layer of the capsule network itself, a facial expression recognition method combining improved ResNet18 and capsule network is proposed. This method only retains the convolutional layer of ResNet18, adjusts the three residual blocks and incorporates the CBAM attention mechanism, then replaces the single convolutional layer of the capsule network to extract features, and finally sends the extracted features to the capsule network for training and classification. The proposed method was tested on the three datasets which recognition accuracies are 9797% on the CK+, 8411% on the RAF-db, 8624% on the FER+ (single label), and 9414% on the FER+ (double label). Experimental results show that the method is feasible and effective in facial expression recognition.

Key words: capsule network; ResNet18; face expression recognition; feature extraction; CBAM attention mechanism

收稿日期： 2022-10-28

基金项目： 国家自然科学基金项目(61971253); 青岛科技大学2021年大学生创新训练计划项目(S202110426006).

作者简介： 刘宁(1995—),男,硕士研究生.*通信联系人.