计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (19): 179-188.DOI: 10.3778/j.issn.1002-8331.2102-0296

• 模式识别与人工智能 • 上一篇    下一篇

利用优化剪枝GoogLeNet的人脸表情识别方法

张宏丽,白翔宇   

  1. 1.内蒙古师范大学 教育技术系,呼和浩特 010022
    2.内蒙古大学 计算机学院,呼和浩特 010021
  • 出版日期:2021-10-01 发布日期:2021-09-29

Facial Expression Recognition Method Using Optimized Pruning GoogLeNet

ZHANG Hongli, BAI Xiangyu   

  1. 1.Department of Educational Technology, Inner Mongolia Normal University, Huhehot 010022, China
    2.College of Computer Science, Inner Mongolia University, Huhehot 010021, China
  • Online:2021-10-01 Published:2021-09-29

摘要:

为了提高人脸表情识别的准确率和加快处理速度,提出了一种基于优化剪枝GoogLeNet的人脸表情识别方法。利用GoogLeNet网络提取面部特征,其中Inception模块加深学习深度,并利用典型的分类器实现人脸表情分类。改进GoogLeNet网络,添加全局最大池化层并保留检测目标的位置信息,以Sigmoid交叉熵作为训练目标,获得全面的人脸表情特征信息。通过剪枝算法对GoogLeNet网络进行训练、修剪低权重连接和再训练网络等操作,以简化网络结构和参数量,提高运行效率。在JAFFE、CK+和Cohn-Kanade数据集上对所提方法进行验证,实验结果表明,所提方法的识别准确率分别为83.84%、85.09%和84.87%,运行时间低于200?ms,优于对比方法,具有较好的适用性。

关键词: 剪枝算法, GoogLeNet, 人脸表情识别, Inception模块, 全局最大池化层, 运行效率

Abstract:

In order to improve the accuracy of facial expression recognition and speed up the processing, a facial expression recognition method based on optimized pruning GoogLeNet is proposed. Firstly, the GoogLeNet network is used to extract facial features, in which the concept module can deepen the learning depth, and the typical classifier is used to realize facial expression classification. Then, the global maximum pooling layer is added to improve the GoogLeNet network, and the location information of the detected target is retained. The sigmoid cross entropy is used as the training target to obtain the comprehensive facial expression feature information. Finally, the pruning algorithm is used to train the GoogLeNet network, pruning the low weight connection and retraining the network, so as to simplify the network structure and parameters and improve the operation efficiency. The proposed method is tested on the data sets of JAFFE, CK+ and Cohn Kanade. The results show that the recognition accuracy of the proposed method is 83.84%, 85.09% and 84.87% respectively, and the running time is less than 200?ms, which is better than the comparison method and has certain universality.

Key words: pruning algorithm, GoogLeNet, facial expression recognition, Inception module, global maximum pooling layer, operating efficiency