OQ-FERNet: An occlusion-aware quantitative facial expression recognition network for real-time human-computer interaction
TL;DR
OQ-FERNet achieves competitive classification performance, improved robustness under different occlusion conditions, and effective quantitative emotion prediction, indicating that OQ-FERNet provides an effective and efficient solution for lightweight, occlusion-robust, and fine-grained facial expression recognition in real-time human-computer interaction scenarios.
Abstract
Facial expression recognition is an important task in real-time human-computer interaction, but its practical application is still limited by partial facial occlusion, high computational cost, and the lack of fine-grained emotion representation. To address these issues, this paper proposes OQ-FERNet, an occlusion-aware quantitative facial expression recognition network. The proposed network adopts MobileNetV4-Conv-S as a lightweight facial expression feature extraction backbone to obtain compact and discriminative representations with low computational cost. On this basis, an Occlusion-aware Regional Reweighting module is designed to divide the feature map into upper-face, middle-face, and lower-face regions, and adaptively adjust their contributions according to regional reliability. This design suppresses occluded or less informative facial areas while enhancing visible expression-related cues. Furthermore, a Quantitative Emotion Prediction Head is introduced to jointly perform discrete expression classification and emotion intensity estimation, enabling the model to provide both categorical and fine-grained affective outputs. Experiments are conducted on RAF-DB and AffectNet-7. The results show that OQ-FERNet achieves competitive classification performance, improved robustness under different occlusion conditions, and effective quantitative emotion prediction. Specifically, OQ-FERNet obtains 93.12% accuracy on RAF-DB and 68.32% accuracy on AffectNet-7, while achieving an MAE of 0.246, an MSE of 0.101, and an RMSE of 0.318 for quantitative emotion prediction on AffectNet-7. In addition, the model contains only 4.18M parameters and 0.26G FLOPs, with an inference speed of 356 FPS. These results indicate that OQ-FERNet provides an effective and efficient solution for lightweight, occlusion-robust, and fine-grained facial expression recognition in real-time human-computer interaction scenarios. The source code is publicly available at: https://github.com/jiahao001-j/OQ-FERNet.