Neonatal Behavioral Monitoring Dataset: Face-Cropped Images from NICU for preterm babies classification

This dataset consists of face-cropped images of neonates captured in a Neonatal Intensive Care Unit (NICU) environment, designed to support vision-based behavioral monitoring research. It focuses on two primary binary classification tasks: distinguishing between sleep and awake states, and identifying crying versus normal (non-crying) behaviors. The images were curated from publicly available sources, emphasizing real-world NICU conditions such as variable lighting, occlusions, and subtle facial cues relevant to low-resource settings.
The dataset is structured into two subsets:

Sleep/Awake Subset: Contains 1,336 images labeled based on visual criteria.
Sleep: 733 images (characterized by eyes closed and minimal movement).
Awake: 603 images (characterized by eyes open and visible activity).

Crying/Normal Subset: Contains 4,338 images labeled using indicators like mouth openness, facial strain, and contextual cues.
Crying: 1,971 images.
Normal: 2,367 images.

All images are preprocessed as face-cropped to focus on relevant facial features, facilitating efficient model training for edge devices. This dataset was utilized in the development and evaluation of BabyEHANet, a hybrid attention network for real-time neonatal monitoring, as described in the associated paper: "BabyEHANet: Dual-Residual Hybrid Attention for Real-Time Vision-Based Neonatal Behavioral Monitoring on Edge Devices."
The dataset is intended for researchers in computer vision, neonatal care, and edge computing, enabling advancements in automated distress detection and behavioral analysis. It promotes accessibility in resource-constrained environments by providing a benchmark for models optimized for hardware like Raspberry Pi.

本数据集包含新生儿重症监护室（Neonatal Intensive Care Unit, NICU）环境中采集的人脸裁剪图像，旨在支撑基于视觉的行为监测研究。其聚焦两项核心二分类任务：一是区分睡眠与清醒状态，二是识别啼哭与正常（非啼哭）行为。本数据集的图像均从公开数据源中遴选得到，重点贴合真实NICU场景的特征，包括可变光照、遮挡情况，以及适用于低资源环境的细微面部特征。

数据集分为两个子集：
睡眠/清醒子集：共1336张图像，基于视觉判定标准进行标注。其中睡眠状态图像733张，特征为双眼闭合且活动幅度极小；清醒状态图像603张，特征为双眼睁开且存在可见活动。
啼哭/正常子集：共4338张图像，基于嘴巴张开程度、面部紧张度及上下文线索进行标注。其中啼哭状态图像1971张，正常状态图像2367张。

所有图像均经过人脸裁剪预处理，以聚焦关键面部特征，可助力边缘设备高效开展模型训练。本数据集曾用于开发与评估BabyEHANet——一种面向实时新生儿监测的混合注意力网络，相关研究论文为"BabyEHANet: Dual-Residual Hybrid Attention for Real-Time Vision-Based Neonatal Behavioral Monitoring on Edge Devices."

本数据集面向计算机视觉、新生儿护理与边缘计算领域的研究人员，可为自动化应激检测与行为分析领域的技术进步提供支撑。其通过为树莓派等硬件的优化模型提供基准，提升资源受限环境下相关研究的可及性。