Video Amplification and Deep Learning in Micro-Expression Recognition

To address the difficulty in identification of micro-expressions and low practical significance of consolidated emotional categories,a video amplification method based on eye interference elimination is proposed,and convolutional neural network(CNN)is used to realize micro-expression recognition.First,we amplify the video data of CASME,CASME II using phase-based video motion processing technology.Then,to eliminate eye interference,feature point location is used to obtain eye coordinates,and replace the original eye video into an enlarged video with fusion processing.Finally,the idea of VGG16 is used to construct CNN model,and identify emotion categories in the enlarged micro-expression data.Experiments compared the accuracy of two datasets under different methods as well as the accuracy of original and enlarged datasets under several models with different tuning strategies.The results show that the method can effectively improve the recognition accuracy under real emotion categories.