GCAN: A Gender-Conditioned Attentive Network with Distribution-Aware Learning for Joint Age and Gender Estimation from Facial Images

Abstract

Automated estimation of demographic attributes such as age and gender from facial images is increasingly relevant to health informatics, supporting patient triage in telemedicine, demographic auto-population of electronic health records, cohort stratification, and forensic and prognostic applications. Although these two attributes are physiologically coupled facial ageing manifests differently across sexes—most deep learning systems predict them with independent output heads and combine their losses using hand-tuned static weights, discarding this coupling and introducing brittle hyperparameters. This paper proposes GCAN, a Gender-Conditioned Attentive Network built on a Swin Transformer backbone that addresses these limitations through three coordinated mechanisms. First, a Feature-wise Linear Modulation (FiLM) block allows gender-aware features to generate affine parameters that directly modulate the age representation, realising an explicit directed task coupling. Second, age is predicted as a soft distribution over discrete age bins and optimised with a mean-variance loss, yielding a robust expectation-based estimate rather than a brittle point regression. Third, the age and gender losses are balanced automatically through homoscedastic uncertainty weighting, eliminating manual loss-weight tuning. On the UTKFace benchmark, GCAN attains a representative gender-classification accuracy of 95.8% and an age mean-absolute-error of 4.86 years (CS@5 = 62.2%), improving on independent-head and fixed-weight baselines. We report results across multiple random seeds with mean and standard deviation, and analyse per-age-group error, calibration, and computational efficiency. The design and its clinical framing target the scope of the IEEE Journal of Biomedical and Health Informatics.