Vec2UAge: Enhancing underage age estimation performance through facial embeddings. (April 2021)

• Record Type: Journal Article
• Authors: Anda, Felix; Dixon, Edward; Bou-Harb, Elias; Le-Khac, Nhien-An; Scanlon, Mark
• Is Part Of: Forensic science international. Volume 36(2021)Supplement
• Publication Date: 2021-04
• Subjects: Child Sexual Exploitation/Abuse Material (CSEM/CSAM) -- Age estimation -- Underage facial age -- Deep learning
• DOI: 10.1016/j.fsidi.2021.301119
• Languages: English
• ISSNs: 2666-2817

Abstract

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