Image based Facial Micro Expression Recognition using Attention Residual Network on Standard and Real time Datasets

Abstract

In Human Computer Interaction (HCI), more researchers are interested in developing newlinereal-time application, especially Micro expression Recognition (MERs) of humans. newlineDifferent methodologies are used to develop the models to recognize the MER. However, newlinestill there are scope to improve the performances and consistency. Our motive is newlineto develop the efficient trained models of MERs for improving the performances and newlinemake it more consistent. We propose, two different models for MERs and an implementation newlineto a real time problem.. The Model-1 is Attention Residual Network For newlineMicro-Expression Using Image Analysis, Model-2 is the Residual Attention Network newlineFor Deep Face Recognition Using Micro-Expression Image Analysis and implementation newlineof model 2 is to Classify High And Low Level ASD Kids Using Attention Mechanism newlineEmbedded Deep Learning Technique. These three proposed models are applied on newlinethe 3 standard and a real time database to validate the MERs. Model-1 is based on ABN newlineidentification, the our contribution is to identify the seven expressions to form the ABN newlinewhich contributes for the MER. An efficient trained model is constructed by applying newlinethe Softmax function with cross-validation in the supervised classifier. The drawback newlineof this model is Vanishing Gradient problem which reduces the performance. Model-2 newlineis developed to overcome this problem. The novelty of the Model-2 work is to identify newlinethe most effective ways to resolve the vanishing gradient problem with ResNet s newlinethrough, which basic seven expressions are recognized. ResNet is chosen based on the newlinelower time complexity and have lower error ( 6.7% at 50 layer), having skip connections newlineallows the network to more easily learn identity-like mappings. The model-2 is newlinemore f1-score and less computation time compared to model 1. The success of Model- newline2 is implemented in the real time application, which is an efficient trained model to newlineclassify High And low level ASD Kids and this is our third contribution. In the experimental newlineresults, the overall validation performance,