InfantFace: Detecting infant faces in neonatal clinical environments

Reliable face detection is a prerequisite for non-contact clinical monitoring, such as automated pain scoring and cardiorespiratory signal extraction. However, neonatal clinical environments present unique challenges, including medical device obstructions (e.g., feeding tubes, goggles), variable lighting, and cluttered backgrounds. This study aims to develop a robust, end-to-end face detection pipeline specifically tailored for infants in these high-acuity settings.

The researchers utilized a one-stage YOLOv11m architecture, a state-of-the-art object detection model. The development process involved a multi-stage training strategy: initial pre-training on large-scale public datasets (VGGFace2, CelebA, FDDB, and WIDER FACE) followed by fine-tuning on a proprietary neonatal research dataset. This dataset consisted of 228 video recordings from 113 infants undergoing routine clinical procedures. The authors compared their model against three established general-purpose detectors (MTCNN, SCRFD, and DSFD) using standard object detection metrics, specifically Average Precision (AP) at an Intersection over Union (IoU) threshold of 0.5 (AP50) and the stricter mean AP across thresholds (AP50:95).

Before clinical domain adaptation, the InfantFace model already outperformed the baseline general detectors, achieving an AP50 of 0.87. After fine-tuning on the neonatal research dataset, performance improved to an AP50 of 0.96. The study demonstrates that while general detectors struggle with the specific visual noise of neonatal units, targeted domain adaptation significantly enhances both detection reliability and localization precision. The authors also noted that models fine-tuned on the WIDER FACE dataset generalized better to the neonatal domain than those trained on other public datasets, suggesting that visual similarity in training data is critical for performance in specialized clinical environments.