Facial Expression Recognition (FER) is a significant task in human–computer interaction, affective computing, and behavior analysis. Conventional methods rely on handcrafted features, which often fail to generalize over variations in illumination, pose, and occlusion. Deep learning, particularly convolutional neural networks (CNNs) and transformer-based architectures, has dramatically advanced FER performance by automatically learning hierarchical features from raw image data. This paper explores recent developments in deep learning–based FER, including CNNs, recurrent neural networks (RNNs), attention mechanisms, and multimodal approaches. We discuss the impact of large-scale labeled datasets, data augmentation, and domain adaptation techniques on improving model robustness. Experimental results demonstrate that deep learning models outperform traditional methods, achieving state-of-the-art accuracy on benchmark datasets such as FER-2013, CK+, and RAF. Despite these advances, challenges remain in handling real-world variability, class imbalance, and interpretability. Future directions include leveraging self-supervised learning, few-shot learning, and hybrid architectures to further enhance FER performance and generalization.