Destruction and Reconstruction Learning for Facial Expression Recognition

Abstract

The most discriminative expression features are mostly concentrated in local key facial regions. Thus, we propose a simple and efficient framework that can learn more discriminative expression features from scrambled facial images. Specifically, we first divide the input image into local subregions of the same size and shuffle them randomly at a certain range to obtain the damaged image to increase the difficulty of recognition. Then, the original image and the damaged image are fed to the network. A channel attention module is exploited for highlighting the effective features and suppressing irrelevant features. Simultaneously, during the reconstruction phase, a region alignment model is appended to establish the semantic correlation between each subregion, aiming at restoring the original spatial layout of local subregions in the original image. Extensive experiments on the RAF-DB and the FERPlus datasets demonstrate that our proposed method significantly outperformed state-of-the-art methods without any external facial expression pretraining.