Learning Framework for Compound Facial Emotion Recognition

Background: Facial emotion recognition (FER) is a vital research area in machine vision and artificial intelligence due to its application in academics and industry. Although FER can primarily be conducted using multiple sensors, research shows that using facial images/videos to recognize facial expressions is a better way to convey emotions because visual expressions carry essential information. Objective: This paper focuses on implementing learning frameworks that combine machine learning and deep learning for detecting 50 classes of compound emotions using the iCV Multi- Emotion Facial Expression Dataset (iCV-MEFED). Methods: In the proposed methodology, we used a deep learning Inception v3 CNN-based model to extract features for each image, and a Multi-Class Support Vector Machine (mSVM) classifier was used to detect the corresponding 50 classes of basic and compound emotions. Results: The proposed learning framework for the iCV-MEFED dataset has an accuracy of 26%, outperforming the state-of-the-art results. Conclusion: Moreover, the results got are compared with competition results in terms of misclassification results, which shows our methodology gives the best result of 74.00%.