Emotion Recognition on FER-2013: A Deep Learning-Based Comparative Analysis

Recognizing human emotions through facial expressions plays a vital role in improving human-machine interaction and smart monitoring systems. This study compares the performance of three models-ResNet50, DeepFace, and a custom CNN-on the FER2013 dataset. The experimental results show that pre-trained models such as DeepFace and ResNet50 significantly outperform the custom CNN, confirming the effectiveness of transfer learning and deep hierarchical feature representations for emotion recognition tasks. Models were evaluated using precision, recall, F1-score, and accuracy metrics, and consistent preprocessing ensured fair comparisons across architectures.