Emotion Recognition using Micro-expressions

Abstract

Micro-expressions (MEs) are involuntary, subtle expressions which can reveal concealed emotions that people don’t want to show. However, analyzing such rapid facial micro-expressions is very challenging due to their short duration and low intensity. Here, we are emphasizing on macro & micro-expressions recognition on diverse Indian faces and emotions. There are biases in the result due to lack of diversity in the available datasets i.e there are only one or two types of facial features, skin tones, etc. included in the dataset. This leads to misleading results and do not recognize varied real time input. The given macro-expression & micro-expression datasets are cleaned and pre-processed. Pre-processing includes noise removal, cropping and conversion of images to grayscale followed by segmentation. The action units in a large macro-expression dataset is tested and designed to map the data with various macro-expressions followed by training the weights on the provided dataset of micro-expressions using transfer learning. The model is then trained using deep Convolutional neural layers obtaining validation accuracy of 76.9% for macro-expressions and accuracy 71% for micro-expressions respectively which is better than other techniques using CNN.