SynChildFace: Fine-tuning Face Recognition for Children with Synthetic Data

Many forensic investigations (e.g., the search for missing children or analysis of child pornographic data) require robust face recognition of children. However, face recognition systems are primarily designed for recognizing adults, and previous studies have already shown that they do not reliably recognize children. Moreover, the limited availability of public datasets and ethical as well as privacy concerns pose challenges for the development of specialized face recognition for children.This work investigates the feasibility of fine-tuning state-of-the-art face recognition models for the task of recognizing children using synthetic face images. Specifically, the public HDA-SynChildFaces database is employed for fine-tuning the MagFace and AdaFace open-source face recognition models. In comprehensive evaluations on the real Young Labelled Faces in the Wild database, significant reductions in terms of error rates are obtained, which confirms the soundness of the proposed approach.

I. Introduction

Face recognition technologies are increasingly being used in forensic investigations [25]. In the past years, face recognition error rates dropped massively, particularly through the use of deep neural networks. However, deep neural networks require large amounts of annotated training data in order to achieve a high recognition accuracy. In addition, it is essential that the training data used reflects the subsequent use case (i.e., domain) as realistically as possible. For example, facial recognition algorithms for access control in critical infrastructures should be trained on facial images of adult subjects. In contrast, some forensic investigations could benefit from algorithms that are able to automatically process large quantities of image material with the goal of identifying children. However, major challenges arise towards the development of specialized face recognition for children: due to data protection reasons, image data that has been seized for criminal prosecution may not generally be shared with algorithm developers and research institutions. In addition, such data would require timeconsuming manual annotation, since identity labels may not be available.