Co-regularized Facial Age Estimation with Graph-Causal Learning

In this paper, we present a graph-causal regularization (GCR) for robust facial age estimation. Existing label facial age estimation methods often suffer from overfitting and overconfidence issues due to limited data and domain bias. To address these challenges and leveraging the chronological correlation of age labels, we propose a dynamic graph learning method that enforces causal regularization to discover an attentive feature space while preserving age label dependencies. To mitigate domain bias and enhance aging details, our approach incorporates counterfactual attention and bilateral pooling fusion techniques. Consequently, the proposed GCR achieves reliable feature learning and accurate ordinal decision-making within a globally-tuned framework. Extensive experiments under widely-used protocols demonstrate the superior performance of GCR compared to state-of-the-art approaches.