Human Facial Emotion Recognition Using Multi-head Cross Attention Network and Attention Consistency

Abstract

In this report we discuss some methodologies related to facial expression recog nition(FER). Facial expressions can be recognized with the help of collective representation of multiple facial regions and proper decoding of the high-order interactions between the local features is very necessary to recognize a particu lar expression efficiently. However, if noise or inconsistency is present, the FER task becomes error-prone. Because of the noise involvement in the samples, the performance of a model degrades. So, it becomes compulsory to cope with the inconsistency at first. Thus we present a model which will try to recognize facial expressions with the ability to focus on multiple regions and tackle the noise in volvement or annotation ambiguity by suppressing it’s effect during the training