基于LBP与双时空神经网络的微表情识别

Micro-expression Recognition Based on LBP and Two-stream Spatial-temporal Neural Network

针对传统微表情识别方法识别率低及过程复杂等问题,设计了一种浅层的双时空多尺度神经网络TSTNet (Two-Stream spatial-Temporal Network)模型.利用局部二值模式(LBP)提取SMIC和CASMEⅡ微表情数据库的纹理特性,将其输入到组合的3维卷积神经网络(3DCNN)与卷积的长短期记忆网络(ConvLSTM)中同时提取时间和空间信息,在模型中加入丢弃算法并多路提取特征,减小过拟合风险的同时学习更丰富的特征.在SMIC和CASMEⅡ微表情数据库上的识别率分别达到了67.30%和65.34%,与现有的深度学习方法相比,该模型提高了网络的训练速度与微表情的识别率.

In this study,we design a two-stream spatial-temporal network model to solve the problems of low recognition rate and complex processes in traditional micro-expression recognition methods. We use a local binary pattern to extract texture characteristics from the SMIC and CASMEⅡ micro-expression databases,and input them into the combined 3D convolutional neural network and convolutional long short-term memory to extract time and spatial information simultaneously. We add a discard algorithm to the model to enable the extraction of multiple features to reduce the risk of overfitting while learning richer features. In the SMIC and CASMEⅡ micro-expressions databases,our recognition rate reached 67.30% and 65.34%,respectively. Compared with existing recognition methods,the proposed model improves the network training speed and the micro-expression recognition rate.