LBAN-IL: A novel method of high discriminative representation for facial expression recognition

Abstract

Existing facial expression recognition (FER) works have achieved significant progress on constrained datasets. However, these methods only consider the sample distribution and achieve limited performance on unconstrained datasets. Facial expressions in the wild are influenced by various factors, e.g. illumination and partial occlusion, providing great challenge for model design and putting forward the higher requirement for feature discrimination. In this paper, we propose a novel LBAN-IL for FER in the wild, including local binary attention network (LBAN) and islets loss (IL). LBAN is based on two operations, local binary standard layer and encoder-decoder module. The former is derived from local binary convolution, so as to prevent excessive sparseness of feature maps and reduce the number of learnable parameters. The purpose of the latter is to generate attention-aware features and accurately discover local changes in the face. The proposed IL aims to enhance the discrimination of expression features by increasing the amplitude of vectors. Experimental results on RAF-DB, SFEW 2.0, FER-2013 and ExpW datasets validate the effectiveness of LBAN-IL and perform over some state-of-the-art methods.