<p>Recent deep learning models have demonstrated outstanding performance in age estimation. However, these models still struggle to distinguish neighboring ages, especially in the elderly range, where the performance drops sharply. Elderly samples are scarce, and their wrinkle features, as high-frequency components, are easily aliased into lower frequencies through repeated downsampling. To address this problem, we propose a dual-branch architecture comprising a global appearance branch and a texture enhancement branch. For the texture branch, a facial texture map is first derived using orientation-sensitive filter banks, and then a CNN encoder extracts high-frequency texture features on this map. These high-frequency features are then bidirectionally fused with multi-scale global appearance features. Besides, we treat each fused feature as a Gaussian distribution in a probabilistic embedding space. Usually, these distributions are subject to ordinal distribution constraints, and the triplet loss is a common approach to enforce the ordinal nature of age. However, the traditional triplet loss fails to reflect age differences in terms of inter-sample distances. Therefore, we propose an adaptive triplet, where a dynamic margin adaptively adjusts to varying age differences. The proposed HEAT method is evaluated on five datasets (MORPH II, CACD, AFAD, UTKFace, FG-NET) and achieves superior performance compared to state-of-the-art methods, with notable improvements in elderly age estimation. By integrating a dual-branch architecture and adaptive triplet Loss, HEAT effectively enhances age estimation accuracy–especially in distinguishing adjacent elderly ages–addressing key limitations of existing deep learning models in elderly age estimation.</p>