Vec2UAge: Enhancing underage age estimation performance through facial embeddings

摘要: Abstract Automated facial age estimation has drawn increasing attention in recent years. Several applications relevant to digital forensic investigations include the identification of victims, suspects and missing children, decrease investigators’ exposure psychologically impacting material. Nevertheless, due lack accurately labelled datasets, particularly for underage range, sufficient performance accuracy remains a major challenge field estimation. To address problem, novel regression-based model was created, Vec2UAge. FaceNet embeddings were extracted used as feature vectors train from VisAGe Selfie-FV datasets. A balanced, unbiased dataset created testing validation. Data augmentation techniques evaluated further be expand training dataset. The learning rate (lr) is one most important hyper-parameters deep neural networks; cyclic approach find optimal initial value lr evaluated. distribution presented per optimiser winning models with Stochastic Weight Averaging (SWA) optimised run reached mean absolute error low 2.36 Additionally, time convergence using SWA significantly faster than other optimisers evaluated, i.e., ADAGRAD, ADAM Gradient Descent. evaluation metric form rather single value, giving more insights into effects random initialisations, on outcome.