基于ACNN和Bi-LSTM的微表情识别

Micro-Expression Recognition Based on ACNN and Bi-LSTM

针对微表情动作幅度小、强度低等缺点,提出了一种基于带有注意力机制的卷积神经网络(ACNN)和双向长短期记忆网络(Bi-LSTM)相结合的神经网络结构。实验采用CASME II数据集,为了减少出现过拟合的风险,首先将预处理后的特征向量经过预训练的VGG16网络提取出基本特征,接着对输出特征进行裁剪,得到带有局部特征的24个微表情识别块和带有整个图片特征的全局特征向量;然后将24个识别块分别经过局部识别块注意力卷积神经网络(BR-ACNN)提取出带有注意力信息的局部特征,将全局特征向量经过全局注意力卷积神经网络(GR-ACNN)提取出带有注意力信息的全局特征;最后,将提取的局部和全局特征,经过Bi-LSTM提取出微表情序列之间的相关性信息。实验结果显示,5折交叉验证平均准确率为0.69,UF_(1)为0.638 2,UAR为0.675 0。CASME II数据集上结果显示,所提算法模型相对OFFApexNet模型,其UF_(1)提高了0.028 1,UAR提高了0.096 9;相对ATNet模型,其UF_(1)提高了0.007 2,UAR提高了0.032 0。

In view of the flaws of micro-expression characterized with a small amplitude and low intensity,a neural network structure has thus been proposed based on the combination of convolution neural network with attention mechanism(ACNN)and bi-directional long short-term memory(Bi-LSTM).CASME II data set has been adopted in the experiment so as to reduce the risk of over-fitting,with the basic features extracted from the preprocessed feature vectors through the pre-trained VGG16 network,followed by the cropping of the output features,thus obtaining 24 micro-expression recognition blocks with local features and global feature vectors with the whole picture features.Next,based on an extraction of local features with attention information from 24 recognition blocks through local recognition block attention convolution neural network(BR-ACNN),global features with attention information are to be extracted as well from global feature vectors through global attention convolution neural network(GR-ACNN).Finally,the correlation information between the micro expression sequences can be extracted by Bi-LSTM based on the extracted local and global features.The experimental results show that the average accuracy rate of 5-fold cross validation is 0.69,UF_(1) is 0.6382,and UAR is 0.6750.The results on the CASME II data set show that the proposed algorithm model,compared with OFFApexNet model,is 0.0281 higher in UF_(1),and 0.0969 higher in UAR;while compared with ATNet model,it has increased by 0.0072 in UF_(1) and by 0.0320 in UAR.