Advanced Facial Expression Recognition Using Model Averaging Ensembles of Convolutional Neural Networks and CAM Analysis

Facial expressions play a vital role in non-verbal communication, conveying a wide range of emotions and messages. Although prior research achieved notable advances through architecture design or dataset-specific optimization, few studies have integrated multiple advanced techniques into a unified facial expression recognition (FER) pipeline. Addressing this gap, we propose a comprehensive approach that combines (i) multiple pre-trained CNNs, (ii) MTCNN-based face detection for improved facial region localization, and (iii) Grad-CAM-based interpretability. While MTCNN enhances the quality of face localization, it may slightly affect classification accuracy by focusing on cleaner yet more challenging samples. We evaluate four pre-trained models - DenseNet121, ResNet-50, ResNet18, and MobileNetV2 - on two datasets: Raf-DB and Cleaned-FER2013. The proposed pipeline demonstrates consistent improvements in interpretability and overall system robustness. The results emphasize the strength of integrating face detection, transfer learning, and interpretability techniques within a single framework can significantly enhance the transparency and reliability of FER systems. Combining FER with EEG-based systems significantly enhances the emotional intelligence of brain-computer interfaces, enabling more adaptive and personalized user experiences. With this approach the paper bridges the gap between affective computing and cognitive neuroscience, aligning closely EEG-centered interaction methodologies. Besides understanding the relationship between facial expressions of emotions and EEG signals will be an important study for literature.