Facial Expression Recognition Dataset with 1.6 Million Balanced Images

FER2025 is a dataset of 1,589,810 facial images annotated with one of seven emotion classes: Angry, Disgust, Fear, Happy, Neutral, Sad, and Surprise. The dataset was created by imadhavan and was last updated in September 2025. It is designed for deep learning and computer vision research.

Use Cases

• Train a convolutional neural network to classify the 'Happy' emotion from facial images.
• Benchmark model robustness across balanced classes like 'Disgust' and 'Fear' using the provided image counts.
• Develop a multi-label classifier or explore feature extraction for the seven distinct emotion labels.
• Analyze potential biases in expression recognition by evaluating performance on the 'Neutral' class versus other emotions.

Strengths

• Contains 1,589,810 total images, providing a substantial volume for model training.
• The dataset is balanced, with a maximum class difference of approximately 1.3% across the seven emotion categories.
• Includes 7 clearly defined emotion classes (Angry, Disgust, Fear, Happy, Neutral, Sad, Surprise).

Limitations

• The specific source, collection methodology, and demographic or geographic diversity of the subjects are unknown.
• Image formats are listed but specific image resolutions, quality, and preprocessing details are not provided.

Provenance

Source

huggingface user imadhavan

Freshness

Last updated on 2025-09-18.