Deep Facial Expression Recognition in the Wild

Abstract

Over the last few years, deep learning (DL) based methods have made huge newlinestrides in the field of computer vision. Though performance has dramatically improved in tasks like face detection and face recognition (FR), near human level performance is yet to be achieved in facial expression recognition (FER). This is due newlineto the challenges including variations in pose, variations in illumination, presence newlineof occlusions and presence of noisy annotations. These challenges are generally observed in an uncontrolled environment, also called as in-the-wild scenario. This thesis newlineinvestigates robust FER methods using DL to handle variations in pose, presence newlineof occlusions and noisy annotations. In addition, current state-of-the-art (SOTA) newlineFER models are large in the number of parameters, and consequently memory inefficient and computationally expensive, making them unfit to be deployed in real-time. newlineTo cope with this challenge, we also investigate light-weight models for FER under newlinein-the-wild scenario. newline newline