Hybrid Feature Fusion based model: Facial Expression Recognition in the Wild

Facial Expression Recognition (FER) is an essential task in human–computer interaction and affective computing, with the goal of automatically labeling emotional states from facial expressions. Even with recent improvements based on DL, FER under real-world conditions continues to be challenging because of occlusions, pose variations, illumination changes, and class imbalance—especially for subtle expressions such as fear and disgust. To overcome these shortfalls, we introduce a hybrid feature fusion model that capitalizes on the strengths of the IR-50 and GhostNet backbones. IR-50 extracts high-level identity-aware global features, and GhostNet adds lightweight, efficient, local, fine-grained detail extraction. This two-stream architecture allows our model to generalize more effectively across heterogeneous expression types and visual conditions. We evaluated our method on the RAFDB dataset with an accuracy of $88.59 \%$, better than some of the strong methods like MANet, PACVT, and FaceTopoNet. The model achieves significant improvements in recognizing underrepresented and ambiguous expressions, proving to be effective for robust FER scenarios having partial occlusion, varying lightning situation, different head poses. We also applied XAI techniques such as LIME and SHAP for interpretability of the model.