Bridging the Gap in Facial Age Progression: An Attention Mechanism Approach

Abstract

With the advent of Generative Adversarial Networks (GANs), significant progress has been made in facial aging prediction. However, existing methods still face considerable challenges. Many studies estimate the ages of individuals in images based on their birth dates rather than visual cues, leading to discrepancies between the predicted ages and the actual appearance of facial aging. Moreover, these approaches often overlook racial consistency, resulting in models predominantly tailored to European populations, which limits their generalizability across different races. To address these issues, we propose a novel facial aging prediction framework that employs three independent encoders to model identity, texture features, and facial skeletal structure. We replace traditional convolutional networks with an attention mechanism-based backbone, integrating spatial and channel attention mechanisms to capture both spatial relationships and age-related feature importance. These attention-enhanced feature maps are then processed through a pyramid feature fusion architecture to facilitate multi-scale feature extraction. Our model effectively captures the subtleties of facial aging across different demographics. Extensive experiments and ablation studies demonstrate that our approach excels in preserving identity, ensuring racial consistency, and generating realistic aging effects. The results further highlight the superior ability of the attention mechanisms to extract detailed, localized features essential for facial aging prediction.

Keywords

• Facial aging prediction
• attention mechanism
• image generation
• computer vision