电子测量与仪器学报

胡敏,高永,吴昊,王晓华,黄忠.融合边缘检测和递归神经网络的视频表情识别[J].电子测量与仪器学报,2020,34(7):103-111

融合边缘检测和递归神经网络的视频表情识别

Video facial emotion recognition based on edge detection and recurrent neural network

DOI：

英文关键词:spatio-temporalfeatures edge detection recurrent neural network random search

基金项目:国家自然科学基金(61672202,61673156)、国家自然科学基金-深圳联合基金重点项目(U1613217)资助

作者	单位
胡敏	1. 合肥工业大学计算机与信息学院
高永	1. 合肥工业大学计算机与信息学院
吴昊	1. 合肥工业大学计算机与信息学院
王晓华	1. 合肥工业大学计算机与信息学院
黄忠	2. 安庆师范大学物理与电气工程学院

Author	Institution
Hu Min	1. School of Computer and Information of Hefei University of Technology
Gao Yong	1. School of Computer and Information of Hefei University of Technology
Wu Hao	1. School of Computer and Information of Hefei University of Technology
Wang Xiaohua	1. School of Computer and Information of Hefei University of Technology
Huang Zhong	2. School of Physics and Electronic Engineering, Anqing Normal University

摘要点击次数: 496

全文下载次数: 2153

中文摘要:

为有效解决传统视频人脸表情识别通常只关注单张视频帧的空间特征,而忽略了相邻帧之间隐藏的时间特征的问题, 提出一种结合边缘检测和递归神经网络的视频表情识别方法,利用梯度边缘检测准确地提取输入图像的纹理信息,同时提出一种分片交叉 LSTM 结构,提取出图像序列中隐藏的时空特征。实验在 CK+和 MMI 视频库上进行,在 OCNN-RNN 网络中分别取得 88. 4%和 69. 7%的识别率,在 GCNN-RNN 网络中分别取得 89. 8%和 73. 6%的识别率,最终使用提出的加权随机搜索方法融合 GCNN-RNN 和 OCNN-RNN 两个网络之后,分别取得了 94. 6%和 79. 9%的识别率,均优于单流网络算法,证明了所提算法的有效性。

英文摘要:

In view of the existing algorithms, traditional video emotion-based facial expression recognition method only pays attention to the spatial features of a single video frame, and ignores the hidden temporal features between adjacent frames. Therefore, this paper proposes a novel method to extract features using edge detection and improved recurrent neural network. Gradient edge detection can extract texture information of video frame in a more accurate way,at the same time, a kind of overlapping LSTM structure is proposed, and the recurrent neural network can acquire the hidden spatio-temporal information from the input frames. The experiments in this paper are carried out on the CK + and MMI video database. the result of 88. 4% and 69. 7% are obtained in the OCNN-RNN network respectively, and the outcome of 89. 8% and 73. 6% are acquired in the GCNN-RNN network from each database. and finally the random search is used to weight the fusion of the results of the GCNN-RNN network and the OCNN-RNN network. After the two networks are finally merged, the average recognition rate of the integrated model is 94. 6% and 79. 9% respectively, and the accuracy is better than other algorithms, the effectiveness of the proposed algorithm is proved.

查看全文查看/发表评论下载PDF阅读器