A CNN Approach to Micro-Expressions Detection

Machine Learning and Convolutional Neural Networks (CNN) have significantly increased the performance in image recognition and are being widely adopted to analyze faces based on availability of very large databases of Figure and postures. A hot research interest of the the Face Recognition community is the recognition of different types of facial expressions. Among these, Facial Micro-Expressions (ME's) are of big interest due to subtle movements which can show deep or suppressed emotions of an individual. These micro-expressions are quite prominently being used in security, psychotherapy, neuroscience and other related disciplines. The major challenge encountered while detecting these expressions are their low intensity and short duration. Previous works have used Eulerian Video Magnification (EVM) in conjunction with haar-cascades for face detection which gave misleading results. In this paper, we have proposed a special Convolutional Neural Network (CNN) model for face detection on which EVM is applied for amplifying the micro-expressions to a calculated threshold. Following that, a separately trained CNN is used to classify the formerly detected micro-expression into one of the seven universal micro expressions. Results obtained during the test experiment are presented at the end of the paper.