An Efficient Ensemble Learning Model Integrating Multi‐Branch Sub‐Networks for Facial Expression Recognition

ABSTRACT Accurate facial expression recognition is still challenging due to occlusion and location variability. Reducing computing overhead is also important because facial expression detection systems may be used in real‐time applications. This research provides an effective ensemble learning architecture for facial emotion identification using advanced data augmentation and transfer learning techniques. The architecture uses a multi‐branch sub‐network framework. We chose the EfficientNet‐B0, RegNet_Y_800MF and MobileNetV2 for ensembling because they are significantly smaller in terms of FLOPs and number of parameters than other variations, such as the EfficientNet‐B7 and RegNet_Y_800MF. We included data augmentation methods such as Mixup and CutMix to make our system more resilient to overfitting. As demonstrated by our proposed approach, combining smaller models is more efficient than using a single large model. The proposed architecture achieves state‐of‐the‐art results with an accuracy of 96.42% and 97.55% on the SUFEDB and KDEF datasets, respectively.