ARM: a lightweight module to amend facial expression representation

The classification objects of facial expression recognition (FER) are visual representations based on human faces, so there are certain generic parts between facial expression features that can be exploited to develop a concise feature extraction pattern. In addition, this paper is the first to investigate the flawed nature of padding in the convolutional layer. We find that the accumulation of padding in convolutional networks has an erosive effect on features, resulting in a partial loss of recognition performance. We block the detrimental impacts of padding to some extent at the extraction stage of the representation by secondary processing of the high-dimensional features at the network’s back end, despite the fact that no solution for total padding replacement has been established. Our proposed module named Amend Representation Module (ARM) (1) deconstructs the expression features into face and expression components, making feature extraction easier and feature representation more accurate, (2) weakens the weights of degraded feature pixels and obtains more discriminative representations from eroded features. Experiments on public benchmarks prove that our ARM based on ResNet-18 boosts the performance of FER remarkably.