Analysis of Micro Facial Expression by deep learning methods: Haar, CNN and RNN
Abstract: Facial expressions are due to the actions of the facial muscles located at different facial regions. These expressions are two types: Macro and Micro expressions. The second one is more important in computer vision. Analysis of micro expressions categorized by disgust, happiness, anger, sadness, surprise, contempt, and fear are challenging because of very fast and subtle facial movements. This article presents three deep learning methods: the Haar Cascade Classifier, the Convolution Neural Network (CNN) and the Recurrent Neural Network (RNN) to perform recognition of micro-facial expression analysis. First, Haar Cascade Classifier is used to detect the face as a pre image-processing step. Secondly, those detected faces are passed through series of the Convolutional Neural Network (CNN) layers for the features extraction. Thirdly, the Recurrent Neural Network (RNN) classifies micro facial expressions. Two types of data sets are used for training and testing of the proposed method: Chinese Academy of Sciences Micro-Expression II (CSAME II) and Spontaneous Actions and Micro-Movements (SAMM) database. The test accuracy of SAMM and CASME II are obtained as 84.76%, and 87% respectively. In addition, the distinction between micro facial expressions and non- micro facial expressions are analyzed by the ROC curve.
Keywords: Micro expression, Convolutional Neural Network, Long Short term Memory