Facial Age Estimation Using Hybrid Architecture: Vision Transformers and ResNet50 with Mixup Data Augmentation

Abstract

Facial age classification is an intrinsically hard problem in computer vision, due to the gradual, continuous nature of ageing, the huge inter-subject variability and the visual ambiguity at the boundary of adjacent age classes. Despite recent advances in deep learning, most existing methods are restricted to either convolutional neural networks or transformer-based architectures and cannot simultaneously capture fine-grained local facial texture cues and long-range global contextual dependencies, which are critical for accurate age-class discrimination. In this work, we propose a novel adaptive fusion architecture with dual branches to jointly integrate a pre-trained ResNet50 backbone for local feature extraction and Vision Transformer for global contextual modelling via an Adaptive Feature Fusion Module with a channel-wise attention mechanism.  Our approach is motivated by the observation that age-related changes in faces exist at multiple scales, from fine local texture patterns such as wrinkles and skin degradation, to holistic structural changes across the whole face. To improve generalisation across ambiguous age class boundaries we further augment the training data with Mixup, a technique known to improve generalisation. We performed extensive experiments on three benchmark datasets, MORPH II, UTKFace, and UAGD, each with different characteristics.  Our model achieves an MAE of 3.42 on MORPH II, 4.42 on UTKFace, and 5.63 on UAGD, consistently outperforming classical methods such as OR-CNN, CORAL, Ranking-CNN, and CSCS-Swin. The ablation studies verify that each component of the architecture is important to the final performance. These results show that the complementary integration of convolutional local descriptors and transformer global representations with adaptive attention-based fusion is a robust and generalisable solution for facial age classification on controlled, in-the-wild and uniformly distributed benchmarks.