Deep Learning Based Methods For Micro-expression Recognition Research

Micro-expression is an expression which is quickly changed on a person's face, it is researched by psychology for many years. In recent years, the use of machine learning methods for micro-expression is becoming more and more popular, and it becomes a hot direction. In the past few years of micro-expression recognition research, several teams for the micro-expression study established some data sets for other researchers to use, and put forward some algorithms to solve the problem of micro-expression recognition. Since the 2012 Convolution Neural Network has made a major breakthrough in competition of ImageNet, the deep learning method based on convolution neural network has been getting better results in the field of image recognition. I do in the micro-expression recognition of the main research done in two ways:1.The first method is based on convolution neural network, mainly based on the optical flow of 3-d CNN network structure, the most simple VggNet network with micro-expression video time series of information constitute a three-dimensional information input , And the 3-d optical flow field information in the X and Y directions is processed with the 3-d CNN of the three channels through the 3-d CNN of the original gray scale.Finally, the three-part information is combined and classified.2. The second method is the use of integral projection and LSTM for micro-expression to identify. The existing micro-expression recognition research is mainly based on the local binary model (LBP) improved algorithm and combined with support vector machine (SVM) to identify.Recently, integral projection has been applied in the field of face recognition. The long and short memory network (LSTM), as a cyclic neural network, can be used to process timing data. Therefore, a model combining LSTM-IP with LSTM is proposed, and experiments are carried out in the latest micro-expression database CASME ?. Horizontal and vertical projection vectors are obtained as LSTM inputs and categorized by integral projection, while pre-venting overfitting techniques. The experimental results show that the LSTM-IP algorithm model achieves better accuracy than the previous method.