Abstract:
Recognizing people with their face has received a lot of attention from the research community. However, face recognition for newborns is in a nascent stage. The need for effective identification of newborns has been rising day-by-day, owing to problems including but not limited to infant swapping, abduction, erroneous drug delivery, and even mother-infant mix-ups in hospitals and intensive care units. In many under-developed and developing countries, most of the babies do not have appropriate official documents that can correctly identify them. Thus, there is a need for accurate biometric recognition of newborns and children. In this project, we are exploring two ideas. Firstly, we are analysing the possibility of using face-recognition as a practical and cooperative biometric modality for newborns, i.e., using the face to uniquely identify newborns. A large data-set collected from various publicly available sources is used. Additionally, we are exploring the idea of using multi-modal fusion for biometric recognition of children under the age of 5 years. Until now, several learning algorithms such as convolutional neural networks have been evaluated on our newborn data-set using a specific benchmark protocol. The performance of these algorithms has been evaluated under both forms of face recognition problem - identification as well as verification settings. We have observed that existing algorithms are unable to yield high accuracies on newborn faces. In this semester, we propose a local region based approach which uses autoencoders as feature extractors, and suggest fusion methods for improved results. Further, a database of face images of 128 subjects of ages 2-4 has been collected over two sessions and presented. Baseline experiments using commercial o_-the-shelf (COTS) and open-source systems have been performed.
