Facial expression recognition based on multi-task self-distillation with coarse and fine grained labels

Abstract

Facial expression recognition (FER) plays a crucial role in numerous human–computer interaction systems. For the sake of precise recognition, existing methods often enhance the representational capacity of networks by designing complex network structures or incorporating additional facial information. However, due to redundancy among facial expression features, refining expression-related information to obtain highly discriminative expression features remains challenging. We propose a multi-task self-distillation method with coarse and fine grained labels for FER. To mine the sufficient expression-related information, we construct coarse-grained auxiliary branches that enhance the learning ability of the network based on the prior in the facial expression labels. To map coarse-grained features into a fine-grained feature space, feature alignment modules are then introduced. Then, refined self-distillation is constructed to transfer coarse-grained knowledge to fine-grained features, providing additional guidance for the extraction of discriminative features. Our proposed method achieves state-of-the-art performance on multiple FER benchmarks, demonstrating its superiority.