Local Multi-Head Channel Self-Attention for Facial Expression Recognition

Since the Transformer architecture was introduced in 2017 there has been many\nattempts to bring the self-attention paradigm in the field of computer vision.\nIn this paper we propose a novel self-attention module that can be easily\nintegrated in virtually every convolutional neural network and that is\nspecifically designed for computer vision, the LHC: Local (multi) Head Channel\n(self-attention). LHC is based on two main ideas: first, we think that in\ncomputer vision the best way to leverage the self-attention paradigm is the\nchannel-wise application instead of the more explored spatial attention and\nthat convolution will not be replaced by attention modules like recurrent\nnetworks were in NLP; second, a local approach has the potential to better\novercome the limitations of convolution than global attention. With LHC-Net we\nmanaged to achieve a new state of the art in the famous FER2013 dataset with a\nsignificantly lower complexity and impact on the "host" architecture in terms\nof computational cost when compared with the previous SOTA.\n