Complexity aware center loss for facial expression recognition

Deep metric-based center loss has been widely used to enhance inter-class separability and intra-class compactness of network features and achieved promising results in facial expression recognition (FER) recently. However, existing center loss does not take the complexity of expression samples into consideration, which deteriorates the representativeness of the generated center vectors resulting in suboptimal performance. To solve this problem, we propose a novel complexity aware center loss for FER. Specifically, a multi-category division module is firstly devised to distinguish simple samples and difficult samples for each category based on the entropy value of sample prediction results. Then, an exact representative center module is employed on simple samples to generate a more representative center vector for each category by encouraging greater differences between different categories. Finally, an adaptive distance adjustment module is proposed to reduce the interference of difficult samples in the model learning process to further improve the accuracy of FER by maintaining a suitable distance between difficult samples and their corresponding center vector. Extensive experimental results on two benchmark datasets demonstrate the effectiveness, universality and superiority of our methods. The code will be available at https://github.com/sanjiaobo/CACL.