Multi-Pathway Learning and Competition: Introducing Brain FER Mechanisms to Deep Model

Facial expression recognition (FER) by the human brain is an extraordinarily complex process. Current deep learning-based FER models primarily emulate the abstraction of features by the visual neural system, without adequately considering the unique neural characteristics of human FER. To explore and harness the positive impact of human neural traits and cognitive strategies on deep learning-based FER, and to further integrate brain-inspired mechanisms into machine FER processes, we attempt to improve deep neural network models by drawing on diverse human FER neural and cognitive mechanisms. We incorporated three key mechanisms of human FER into the deep learning models: the fuzzy inter-class boundary, localized region attention, and cognitive differences in expression types. Based on these mechanisms, we proposed optimization strategies including multi-pathway classification, localized perception branches, and competitive decision-making module. In the experimental section, we evaluated the proposed model using widely adopted benchmark datasets and conducted comprehensive comparisons with state-of-the-art methods from various periods. The experimental results demonstrate the effectiveness and superiority of the proposed model, while the incorporation of neural mechanisms also presents a FER pattern more closely aligned with the human brain FER.