Why Face Recognition Struggles With Infants: A Longitudinal Study of Children Aged 0–3 Years

The study by researchers from Clarkson University and the University of North Carolina at Charlotte shows that modern face recognition systems perform very poorly for infants, especially under six months, due to rapid facial changes, but accuracy improves as children grow older. By applying adaptive learning techniques, the researchers demonstrate that long-term recognition performance can be significantly improved, making face recognition more viable for toddlers in healthcare and child safety applications.

Researchers from Clarkson University in New York and the University of North Carolina at Charlotte set out to examine a major blind spot in biometric technology: face recognition for infants and toddlers. While facial recognition systems are widely used for adults, they struggle when applied to very young children whose faces change rapidly in the first years of life. Hospitals, child protection agencies, and smart city systems increasingly depend on reliable identity verification, yet traditional methods such as ID bracelets or ink footprints are error-prone and easily compromised. This study explores whether modern deep learning models can overcome these limitations and provide a dependable solution for identifying children from birth to age three.

Building a Rare Long-Term Face Dataset

To answer this question, the researchers created a unique resource called the Infants and Toddlers Longitudinal Face (ITLF) dataset. It contains 630 facial images of 30 children collected over seven sessions spanning roughly two years. Unlike most existing datasets, ITLF follows the same children over time, capturing how their faces evolve from infancy into early childhood. Images were taken in realistic conditions, often involving uncooperative subjects, natural backgrounds, and varied lighting. While the dataset is relatively small, its strength lies in showing real facial growth patterns rather than idealized studio conditions.

Testing Modern Face Recognition Models

Using the ITLF dataset, the researchers evaluated four popular deep learning face recognition models: FaceNet, ArcFace, CosFace, and MagFace. Each child’s first recorded image was used as an enrollment template, while later images were treated as verification attempts months or years later. When images were compared within the same session, all models performed extremely well, with accuracy above 90 percent. However, this success disappeared once time passed between enrollment and verification. The longer the gap, the worse the systems performed, revealing how sensitive face recognition is to aging in young children.

How Age and Time Affect Accuracy

The results show a clear relationship between age and recognition performance. For infants aged 0 to 6 months, verification accuracy was extremely low. Even the best-performing model correctly verified identities only about 30 percent of the time under strict security conditions. This poor performance reflects how unstable infant facial features are and how similar different infants look at that age. Accuracy gradually improved as children grew older. By the time children reached 2.5 to 3 years, verification accuracy rose to around 65 percent. While still far from perfect, this improvement suggests that facial features become more distinct and stable as toddlers mature. The study also found that short time gaps caused less damage to accuracy than long gaps, confirming that facial change in early childhood is rapid and uneven rather than gradual.

Making Face Recognition More Stable Over Time

To tackle the problem of facial change, the researchers introduced a domain-adversarial learning approach designed to make face representations more stable across time. This method trains the system to focus on identity-related features while ignoring session-specific differences caused by age or growth. When applied to the best-performing model, MagFace, this approach boosted long-term verification accuracy by more than 12 percentage points for older children. The improvement demonstrates that while early childhood facial change is a major obstacle, intelligent model adaptation can significantly reduce its impact.

What This Means for Real-World Use

The study delivers a clear message: face recognition is not reliable for very young infants, especially those under six months old, and should not be used alone for critical identity decisions at that stage. However, it also shows that performance improves with age and can be further strengthened through adaptive learning techniques. These findings are especially relevant for healthcare systems, missing-child investigations, and smart city services that require secure but ethical identification methods. By highlighting both the limits and possibilities of facial biometrics in early childhood, the research lays the groundwork for safer, more age-aware identity systems that respect the biological realities of growing children.