Optimizing Facial Expression Recognition: A One-Class Classification Approach Using ResNet18 and CBAM

Our study addresses the challenge of Facial Expression Recognition (FER) in uncontrolled settings, hindered by imbalanced datasets and insufficient labeled data. We introduce a novel one-class classification method using only positive samples, effectively mitigating dataset imbalances. By integrating the Convolutional Block Attention Module (CBAM) into the ResNet18 network, our approach enhances feature extraction, focusing on crucial facial regions. We initially pre-train on multiclass FER datasets for baseline weights, then employ a One-Class SVM classifier with positive samples, creating expression-specific models. Evaluating on FER2013, AffectNet, and RAFDB datasets, our method shows improved accuracies post-integration, achieving 75.77%, 60.63%, and 89.55% respectively, demonstrating the efficacy of our approach. Our methodology and results are detailed at https://github.com/Accci/OCFER.