Research On Micro-expression Recognition Algorithm Based On Deep Learning

Micro-expression is a subtle and fast facial movement,and can reveal the true emotions that a person tries to conceal.Micro-expression is considered as an important clue of detecting lie,and has attracted the attention of many researchers because of its lots of promising applications in various fields.However,Micro-expression recognition has still faced a huge challenge to researchers for its characteristics of short duration and low intensity in motion,and its accuracy is still low.To improve the recognition accuracy of micro-expressions,in this thesis,we investigate the micro-expression recognition algorithms based on deep learning,and the main work is as follows:(1)A micro-expression recognition method based on hierarchical neural network(HINN)is proposed.Since the movement of the micro-expression of the human face can be composed of the motion state of the local regions of the eyebrows,eyes,nose,and mouth.Firstly,a cascade regression tree algorithm is used to extract 68 feature points for each frame in the micro-expression video,and we divide the micro-expression image in the micro-expression dataset into four sub-images(eyebrows,eyes,nose,and mouth images),then extract rich local features from the sub-images through a layered convolution network,and send the extracted features into a bidirectional recurrent network(BRNN),the BRNN learns the dynamic sequence features,and finally we use fully-connected network to classifies the obtained advanced features.(2)A micro-expression recognition algorithm based on robust principal component analysis(RPCA)and bidirectional LSTM(RBLSTM)is proposed.Firstly,considering the sparse characteristics of micro-expressions,RPCA is used to extract the sparse information of micro-expressions from the image frames of micro-expression videos,and then the extracted sparse information is sent to the bidirectional LSTM to obtain the global sparse features.(3)A micro-expression recognition algorithm based on deep fusion neural network(DFNN)is explored.Linear weighting is used to fuse the HINN network and RBLSTM,and the micro-expressions are classified by combining the extracted sparse-rich and local-global features.The algorithm proposed in this thesis is tested on a data set composed of four spontaneous micro-expression databases(CASME I,CASME II,CAS(ME)2 and SAMM).The experimental results show that the proposed method is compared with existing algorithms has a better performance.