LaurenGurgiolo/Micro_Facial_Expressions

Micro Facial Expressions Dataset

Dataset ID: LaurenGurgiolo/Micro_Facial_Expressions Task: Facial Emotion Recognition License: Refer to source datasets Languages: N/A Domain: Computer Vision, Affective Computing

Dataset Description

The Micro Facial Expressions dataset is a combined facial emotion recognition dataset composed of two sources:

FER-2013 – a large-scale grayscale facial expression dataset

Micro-Expression Image Dataset – a curated collection of color facial images representing subtle emotional expressions

The dataset is designed to support research in emotion classification, facial expression analysis, and affective computing across varying image resolutions, color spaces, and expression intensities.

Emotion Classes

All images are labeled using the same seven emotion categories:

Label Emotion 0 Angry 1 Disgust 2 Fear 3 Happy 4 Sad 5 Surprise 6 Neutral Dataset Components

1. FER-2013 Dataset

The FER-2013 dataset (Sambare, 2020), originally released on Kaggle, consists of:

Image format: Grayscale

Resolution: 48 × 48 pixels

Preprocessing:

Faces automatically aligned

Centered and normalized to occupy a consistent spatial area

Objective: Emotion classification into one of seven classes

Data splits:

Training set: 28,709 images

Public test set: 3,589 images

This dataset provides a robust baseline for training deep learning models on standardized facial emotion recognition tasks.

2. Micro-Expression Image Dataset

The micro-expression dataset was compiled using facial images collected from Google Image Search (Irfan, 2022).

Key characteristics:

Image format: Color (RGB)

Resolution: 80 × 80 pixels

Subjects: Children, adults, and elderly individuals

Emotion categories:

Anger

Disgust

Fear

Happiness

Neutral

Sadness

Surprise

Preprocessing Pipeline

A series of Python scripts were used to ensure dataset quality through the following steps:

Removal of duplicate images

Exclusion of images without detectable faces

Cropping of facial regions

Removal of images smaller than 80 × 80 pixels

Resizing all images to 80 × 80 pixels

Final manual inspection and verification

This preprocessing ensures consistency in facial framing and image quality while preserving subtle emotional cues characteristic of micro-expressions.

Intended Use

This dataset is intended for:

Facial emotion recognition research

Micro-expression analysis

Training and evaluation of deep learning models

Cross-dataset generalization studies (grayscale vs. color, macro vs. micro expressions)

Limitations

Images are sourced from publicly available datasets and web searches, which may introduce demographic or cultural biases.

FER-2013 images are low-resolution grayscale, while micro-expression images are higher-resolution color, potentially affecting model generalization.

Citations

Sambare, M. (2020). FER-2013 Facial Expression Recognition Dataset. Kaggle.

Irfan, M. (2022). Micro-Expression Dataset via Google Image Search.