Emotion Detection Dataset

This dataset contains facial images collected for five-class facial emotion recognition covering the categories anger, fear, happy, neutral, and sad. Images were captured from 50 participants using mobile phone cameras under varying illumination, face orientation, and background conditions to encourage generalisation across realistic acquisition settings. The raw collection comprises 2,050 original images, each manually labelled with a single emotion class and resized to 224 × 224 pixels. To support model training under class imbalance, an augmented version expands the collection to 11,942 images using rotation, horizontal flipping, grayscale conversion, and perspective transformation. The class distribution in the augmented set is sad (2,826), neutral (2,794), happy (2,645), fear (2,447), and anger (1,230). The data is organised into training, validation, and test splits in a 70/15/15 ratio. Splitting was performed at the participant level, so all images from a given subject appear in only one split, preventing identity leakage between sets and enabling subject-independent evaluation. The dataset is suitable for facial emotion recognition, affective computing, human-computer interaction, and transfer learning research, and supports both CNN-based and transformer-based approaches. It was used to evaluate a hybrid EfficientNetV2-S and Vision Transformer architecture for emotion classification.