Facial Emotion Recognition Dataset for Children with Autism

The Facial Emotion Recognition Dataset for Children with Autism (FER-Autism) is a curated and augmented image collection developed to advance research in Autism Spectrum Disorder (ASD) detection and facial emotion recognition using computer vision and deep learning techniques.

The dataset contains 1,200 training images and 220 testing images, carefully selected to ensure diversity and representativeness across emotion categories. It is an enhanced and restructured version of an existing Autism Facial Recognition Dataset. Through systematic data augmentation and a balanced train/test split, this dataset provides a robust foundation for building accurate and generalizable models capable of identifying emotional patterns and ASD-related facial cues in children.

This dataset was developed under the supervision of Prof. Shimaa Elgamal, Lecturer of Neurology, Faculty of Medicine, Kafrelsheikh University (Google Scholar Profile
).

Data Augmentation

To enhance diversity and robustness, a comprehensive data augmentation pipeline was implemented using the Albumentations library. Each original image was augmented 10 times, introducing variations in geometry, color, brightness, and noise to simulate real-world conditions and reduce overfitting.

Key Augmentation Techniques:

Geometric: Horizontal flip and random rotation (±10°).

Color & Brightness: Hue, saturation, and value shifts; gamma contrast adjustment.

Noise & Quality: Gaussian blur and Gaussian noise.

Resizing & Cropping: Random cropping and resizing to 224×224 pixels.

These transformations strengthen the dataset’s capacity to train models that generalize well across different lighting, orientations, and environments.

Emotion Classes

The dataset includes six primary facial emotion categories representing a range of affective states in children:

Natural

Anger

Fear

Joy

Sadness

Surprise
additional help links
https://onlinelibrary.wiley.com/doi/10.1002/aur.70030?utm_source=chatgpt.com
https://arxiv.org/abs/2307.13706?utm_source=chatgpt.com

面向自闭症儿童的面部情绪识别数据集（Facial Emotion Recognition Dataset for Children with Autism，简称FER-Autism）是一套经过精心整理与增强的图像集，旨在借助计算机视觉与深度学习技术，推动自闭症谱系障碍（Autism Spectrum Disorder，简称ASD）检测及面部情绪识别领域的研究进展。

本数据集包含1200张训练图像与220张测试图像，经严格筛选以确保各情绪类别下的样本多样性与代表性。该数据集是对现有自闭症面部识别数据集的增强与重构版本。通过系统化的数据增强策略与均衡的训练/测试集划分，本数据集为构建精准且泛化能力强的模型提供了坚实基础，此类模型可用于识别儿童的情绪模式与自闭症相关面部特征。

本数据集由卡夫尔谢赫大学医学院神经病学讲师希玛·埃尔加马尔（Shimaa Elgamal）教授指导开发（谷歌学术主页）。

数据增强策略
为提升数据集的多样性与鲁棒性，本研究借助Albumentations库构建了一套完整的数据增强流程。每张原始图像均被增强10次，通过引入几何变换、色彩亮度调整、噪声注入等变体，模拟真实世界的拍摄场景，从而降低模型过拟合风险。
核心增强技术包括：
1. 几何变换：水平翻转与随机旋转（±10°）
2. 色彩与亮度调整：色调、饱和度与明度偏移，伽马对比度调节
3. 噪声与画质处理：高斯模糊与高斯噪声注入
4. 裁剪与缩放：随机裁剪并统一调整至224×224像素尺寸
上述变换可有效增强数据集的泛化能力，助力模型在不同光照、拍摄角度与环境下均能实现良好的表现。

情绪类别
本数据集涵盖六类主要面部情绪类别，覆盖儿童的多种情感状态：
自然（Natural）、愤怒（Anger）、恐惧（Fear）、喜悦（Joy）、悲伤（Sadness）、惊讶（Surprise）

辅助参考链接
https://onlinelibrary.wiley.com/doi/10.1002/aur.70030?utm_source=chatgpt.com
https://arxiv.org/abs/2307.13706?utm_source=chatgpt.com