针对公共空间中人脸情绪识别准确率不高的问题, 提出一种结合不同感受野和双流卷积神经网络的人脸情绪识别方法. 首先建立基于公共空间视频的人脸表情数据集; 然后设计一个双流卷积网络, 以尺寸为 224× 224 的单帧人脸图像输入卷积神经网络 (convolution neural network, CNN), 分析图像纹理静态特征; 以尺寸为 336× 336 视频序列输入 CNN 网络, 再将提取的特征送入长短期记忆网络 (long short term memory network, LSTM) 分析局部, 全局运动特征; 最后通过 Softmax 分类器将两通道网络的描述子进行加权融合, 得到分类结果. 结果表明, 本文方法能有效利用不同感受野的信息特征清晰识别公共空间的 4 种典型人脸情绪, 识别准确率达 88.89%.

Abstract: Aiming at the low accuracy of facial emotion recognition in public space, a method of facial emotion recognition, which was based on different receptive fields and two-stream convolution neural network, was proposed. Firstly, a facial expression dataset based on public spatial video was established. Then, a two-stream network was designed, for which a single face image with a size of 224× 224 was input into convolutional neural network (CNN) to analyze the static characteristics of the image, and a 336× 336 video sequence was input to CNN network, the extracted features were then sent to the long and short term memory network (LSTM) to analyze the local and global motion peculiarity. Finally, the softmax classifier was used to fuse the descriptors of the two channel to get the classification results. The results show that this method can effectively identify four typical facial emotions in public space by using the information features of different receptive fields, and the recognition accuracy reaches 88.89%.