Research On Micro-expression Detection Algorithm Based On Spatial And Temporal Features

Micro-expressions are spontaneous facial expressions that occur when people try to cover up their inner emotions,and they can neither be forged nor suppressed.The irrepressible nature of micro-expressions makes it an important basis for judging one's true emotions.Therefore,micro-expressions are widely used in areas such as national security,psycho-medical care,and police law enforcement.The short duration of the micro-expression and the small range of motion make it difficult to manually detect the micro-expression.Therefore,the use of computer technology for micro-expression detection has become a research hotspot in recent years.This article studies the micro-expression detection algorithm,and the specific work is as follows:Because of the current lack of effective micro-expression feature descriptors in the field of micro-expressions,a micro-expression detection method based on feature distance of motion edges feature is proposed in the spatial domain.First,for the problem that the image contains motion noise and blinking noise outside the face range,the ellipse region of interest is extracted to avoid the noise outside the face range.Secondly,aiming at the small movement of micro-expressions,an improved motion edge feature based on the FlowNet2 network is designed.Using the FlowNet2 network can extract the optical flow information of minimal movement,effectively retaining the micro-expression movement information.On this basis,the optical flow gradient amplitude is used to extract the motion edge features,which can effectively remove the head offset error and the cumulative error of the video sequence included in the optical flow field,and has a stronger ability to characterize the micro-expression.Finally,a micro-expression detection method based on multi-frame average feature distance is proposed,and single-frame noise elimination and duration filtering postprocessing are designed to effectively solve the single-frame noise and macro expression sequence noise in the detection process,making micro-expression detection results more accurate.Based on the motion edge feature map,combined with the deep learning algorithm,the MED-TFBL(Micro-expression detection based on temporal features and balanced sample loss)network was designed to realize the micro-expression detection algorithm based on the temporal features and balanced sample loss.First,in order to increase the video sequence samples used for deep learning network training,the sliding window mechanism and the Time Interpolation Model are used to divide and reconstruct the data set,which expands the data volume for training network.Secondly,the motion edge feature map obtained from the algorithm in Chapter 3 is used as the network input.In order to reduce the problem of large data dimension and redundancy of the motion edge feature map,a feature optimization network based on convolutional neural network is designed.Then,a temporal feature extraction network based on Bi LSTM is proposed,and the weight of the temporal features is assigned in conjunction with the attention mechanism,so that the micro-expression temporal features of each frame image correspond to different contribution scores for the microexpression detection.Finally,for the problem of serious imbalance between the number of micro-expression and non-micro-expression samples,a sample balance loss function is proposed to assign weights to the losses of different types of samples,which improves the effect of network training and improves the accuracy of the model.This thesis conducts experiments on the CASME and CASME ? datasets.Experiments show that the micro-expression detection algorithm based on the feature distance of the motion edge feature proposed in this thesis is compared with the current popular traditional algorithm.The test results TPR(True positives rate),FPR(False positives rate),F1(F1Score),ACC(Accuracy)have achieved better results.Compared with the current popular deep learning micro-expression algorithm,the proposed micro-expression detection algorithm based on temporal features and balanced sample loss has achieved higher detection accuracy.