Enhancing Facial Age Estimation with Local and Global Multi-Attention Mechanisms

Recent deep learning models have shown impressive performance in age estimation tasks. However, their effectiveness in real-world scenarios remains limited, especially under challenging conditions such as large pose variations, diverse facial expressions, low image quality, and occlusions. Accurate age estimation necessitates both detailed facial features (e.g., wrinkles, eye bags, laugh lines) and robust global region relationships. Despite this, few existing methods focus on effective feature representations that can identify age-related local facial regions while learning rich global relationships. In this paper, we propose a novel Global and Local Aware (GLA-Age) age estimation framework, comprising three key components: age-related region attention, age Vision Transformer, and adaptive attention dropping strategy. The age-related region attention module captures detailed facial information and identifies discriminative age-related regions. The Regional Vision Transformer enhances local information through long-range and global dependencies, improving robustness in challenging scenarios. Additionally, the Age Adaptive Attention Dropping (A3D) strategy, applied to both the age-related region attention and age Vision Transformer modules, further enhances the diversity and robustness of feature representations. Experimental results demonstrate the superior performance of GLA-Age on widely-used benchmarks such as IMDB-Clean, Morph2, CACD, KANFACE, and FG-NET, under both intra-dataset and cross-dataset evaluation protocols compared with state-of-the-art methods. Furthermore, we create a challenging test set based on the Adience dataset, showcasing its superiority on extremely difficult facial images.