Feature Extraction For Facial Expression Recognition On Video Using Deep Learning

Abstract

The automatic detection and classification of facial expressions from computers have newlinebeen a fascinating topic for research science. Facial expression recognition (FER) has newlinefound its application in many places like Healthcare, Crime, policing, etc. newlineEmotion Recognition can be extremely complex because there is a high correlation newlineamong pixels and frames in terms of spatial and temporal coherence among different newlinecategories of the data. newlineThe fundamental approach behind emotion recognition is the design and effective newlinetraining of machine learning-based algorithms which can detect regular patterns in newlineimage or video data formats and subsequently classify new image/video formats with newlinehigh accuracy. newlineDeep learning is an approach that helps to understand the system for complex tasks with newlinegreater accuracy. A deep learning approach is one of the promising approaches to newlineextracting complex data at a higher level of abstraction. newlineWe propose a framework based on investigating the significance Inception module from newlinethe GoogleNet and InceptionV3 models in grasping facial expression-related newlineinformation. We dub these analyses and create a new module called the YTH (Yielding newlineTensor-based Hierarchy) module. The proposed framework also takes its inspiration newlinefrom the attention network. newlineThe proposed architecture provides significant performance improvements as well as being newlinerelatively simple, easy to train, and easy to implement. We obtain a FER2013 and JAFFE newlinexvii newlinedatabase test accuracy of 70.89% and 92.85%, respectively, outperforming many newlinecomplex architectures. newlineMoreover, we also report accuracy on standard databases to make sure our benchmark newlineis fair against other systems that consume a lot of memory. This model is robust, can be newlinetrained more quickly, and provides high frame-per-second performance. Therefore, a newlinecomplex real-world database can also be handled efficiently.