Facial Emotion Recognition – A Comprehensive Review of Deep Learning and Traditional Learning Approaches with Emerging Challenges

Abstract

In the domains of computer vision and artificial intelligence, facial expression-based emotion recognition (FER) has emerged as a key research area and industrial challenge. Its growing importance is evident in applications such as intelligent surveillance, healthcare monitoring, interactive systems, and mental health assessment. This review systematically examines 95 research papers published between 2014 and 2024, focusing exclusively on studies employing facial imagery for FER. The reviewed works encompass traditional machine learning, deep learning, and hybrid approaches, and analyze their methodologies for preprocessing, data augmentation, and temporal feature extraction. Traditional machine learning methods such as SVM, KNN, and Random Forest achieved average accuracies ranging from 85% to 93%, with the best reaching 99% on the CK+ dataset. Deep learning architectures, including VGG16, ResNet-50, MobileNetV2, and DenseNet-161, reported accuracy improvements of up to 99.5% across benchmark datasets such as FER2013 (35,887 images), CK+, JAFFE, KDEF, and RAF-DB (≈40,000 images). Hybrid models combining CNN feature extraction with classifiers such as SVM and GMM achieved 3–6% improvements in recognition accuracy over standalone CNN or SVM systems, reaching 99.69% on CK+ and 99.53% on JAFFE. By comparing benchmark datasets, algorithms, and performance metrics, this study highlights the significant evolution from handcrafted feature-based systems to fully automated deep learning pipelines. The findings provide a quantitative synthesis of the current state of FER research, offering both newcomers and advanced researchers a comprehensive understanding of the field’s progress, challenges, and future opportunities.