Using VGGFace2 and Long short-term memory network to do the Micro-expression Recognition Task by implementing SMIC datasets.
Step1: Using CNN visualization and GradCam HeatMap technique to visualize the pre-trained VGGFace2 model, which is basically a ResNet-50 CNN networks;
Step2: Pre-processing the data from SMIC micro-expression datastes. In directory: Pre-processing, 4 methods beed used for pre-processing.
@DataPreNoAug:Expending the number of sample videos by fixed-size slide window;
@DataPreAug:Expending the number of sample videos by Random steps between frames within a slide window;
@DataAugFlip:Variabl-size of slide window(if image length smaller than the seleframe, window size = image length); Mirror flipping
@DataAugFlipPad:Fixed-size of slide window(if image length smaller than the seleframe, padding frames); Mirror flipping
Step3: Jointing scracthes from VGGFace2 Model to Bi-LSTM and replace the last classifier layer as SVM model.
Step4:
@Gradcam could be used to visualize the Grad-Cam heat map.