DiffAtten-GAN: A Local Attention-Enhanced Diffusion-GAN for Elderly Facial Expression Synthesis with Limited Data
Year: 2025
Type: conference-paper
Abstract: The rapid aging of the global population has heightened the demand for sophisticated emotion recognition systems in elderly care, yet the limited availability of comprehensive elderly facial expression datasets remains a significant barrier. In this study, we introduce DiffAtten-GAN, a novel generative model that seamlessly integrates a local attention mechanism with a Diffusion-GAN framework to synthesize high-fidelity elderly facial expressions from limited data. Leveraging a diverse dataset of 6,311 images obtained from FairFace [1] and UTKFace [2], our approach effectively captures age-specific features and subtle emotional cues across varied ethnicities. Experimental results demonstrate that DiffAtten-GAN outperforms conventional GAN architectures, achieving a Fréchet Inception Distance (FID) of 10.423 and an Inception Score (IS) of 2.761 -corresponding to improvements of $21.26 \%$ and $10.09 \%$, respectively, over a baseline StyleGAN2 model. This work addresses data scarcity in elderly expression analysis and contributes a novel hybrid architecture, a local attention module, significant FID/IS improvements, and comprehensive evaluations including human and cross-dataset studies.
Authors: Haoyu You, Maziar Nekovee
Institution: University of Sussex
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4415125515