From Faces to Ages: Enhancing Historical Recognition with Transfer Learning

This thesis investigates deep learning-based facial age and gender estimation, focusing on addressing demographic imbalance and enhancing fairness for underrepresented groups such as children and the elderly. Motivated by the practical needs of historical demographic reconstruction and inclusive biometric design, it proposes a modular and interpretable framework that supports both continuous age regression and categorical age-group classification. Five neural architectures are evaluated, ranging from single-task regression to multi-task and cascaded models. All models share a ConvNeXtV2 backbone but differ in task configurations and output strategies. To mitigate the demographic skew in the CityFace dataset (86% male, predominantly aged 19–39), a λ-based balancing strategy is applied, preserving all female and elderly samples while reducing overrepresented youth. Additionally, portraits of Swedish celebrities are incorporated to enrich demographic diversity. Experimental findings show that while balanced training improves overall model stability and group-wise metrics, it also leads to complex subgroup-specific patterns. In particular, residual diagnostics reveal that elderly samples exhibit systematic underestimation and increased prediction variance across regression models (Models 1, 2, and 5). Notably, the cascaded model (Model 5), despite strong general performance, suffers from pronounced degradation in elderly age estimation, highlighting the limitations of naïve demographic balancing in promoting fairness. Subgroup-level MAE, residual plots, and calibration curves are used to expose biases not visible in global metrics. These findings underscore the need to align architecture design, supervision strategy, and fairness objectives in age and gender prediction tasks. In sum, this study provides a reproducible pipeline for fair and interpretable facial attribute estimation on historical portraits, with implications for digital heritage research, forensic profiling, and equitable AI development.