Research On Lie Recognition Method Based On Micro-expression

Traditional lie detection methods,which use skin-to-skin tools such as lie detectors,can make subjects nervous and fearful,making them unable to communicate in natural situations.Studies have shown that micro-expression is an important clue to identify lies.In the face action coding system,any facial expression is composed of action unit.This thesis uses deep learning method to design a non-contact recognition model to complete lie recognition.The main content includes three aspects:the spotting of micro-expression,the recognition of micro-expression action unit and the use of micro-expression action unit to complete lie recognition.The main contributions of this thesis are as follows:Firstly,aiming at the problem of inaccurate positioning of spotting micro-expressions in long videos,a micro-expression spotting algorithm based on attention mechanism and onedimensional convolution sliding window is proposed.Firstly,VGG16 network is used to learn the spatial features of each frame,and BI-LSTM and one-dimensional convolution are used to learn the global features and local features respectively.Then,the attention mechanism is used to highlight the key frame features of micro-expressions because they appear at local positions in the frame sequence.According to the feature of all images,the movement characteristic of one dimension convolution is used to detect the micro-expression of feature interval.Finally,considering that a micro-expression may appear on multiple intervals,the intervals judged as micro-expressions are further combined to obtain the final micro-expression interval.Experiments on CAS(ME)2 data set verify the effectiveness of the proposed algorithm.Secondly,aiming at the problems of low intensity and difficulty in recognizing microexpression action units,a micro-expression action unit recognition algorithm based on dynamic graph and spatial pyramid is proposed.Firstly,the video is passed through the dynamic image generation module to generate a dynamic image,which can extract the motion information contained in all frames.Then,in view of the subtle motion characteristics of micro-expressions,the spatial pyramid is used to obtain the image features of the network from low level to high level.It is also known that micro-expressions occur in local regions of the face,so regional feature network and attention mechanism are used for each layer of image features.Finally,the binary classification model is trained for each action unit to solve the problem of weak correlation between action unit.Experiments on CASME and CAS(ME)2 data sets verify that the algorithm has good action unit recognition performance.Thirdly,aiming at the problem that the image will be affected by light,head movement and other factors,which leads to poor lie recognition effect,a lie recognition algorithm based on micro-expression action unit is proposed.Firstly,the statistics of action unit in all microexpression intervals of videos related to lies obtained by the micro-expression discovery algorithm and the micro-expression action unit recognition algorithm are calculated and formed into one-dimensional feature vectors.Then,a multi-scale feature extraction network is used to extract adjacent features in different ranges.Finally,the multi-head attention mechanism is used to obtain the degree of attention between features.Experiments show that this algorithm can achieve high accuracy on multi-modal lie Dataset of Real-life Trial Data.