Attention-enhanced U-Net with VGG backbone for robust facial wrinkle segmentation under variable illumination and pose conditions

Abstract

Facial wrinkle segmentation is critical for automated dermatological assessment, yet existing deep learning methods exhibit significant performance degradation under real-world illumination and pose variations, restricting practical clinical deployment where imaging conditions cannot be controlled. This study proposes a novel robustness-oriented segmentation framework that integrates three synergistic components: (1) attention-enhanced U-Net architectures with strategically frozen VGG16/VGG19 backbones enabling hierarchical feature transfer, (2) a dual augmentation strategy coupling geometric transformations for pose invariance with a four-level photometric enhancement pipeline for illumination robustness, and (3) a weighted mask fusion mechanism combining expert annotations with weak supervision labels. Three architectures (baseline Attention U-Net, VGG16, and VGG19 variants) are trained on 1,000 FFHQ-Wrinkle images and systematically evaluated across four augmentation strategies under nine challenging deployment conditions, including low light, high contrast, noise, head tilts, and perspective shifts. The proposed VGG19 Attention U-Net with combined augmentation achieves a Dice coefficient of 0.6533 and IoU of 0.4931, outperforming the best existing method (Striped WriNet) by +4.26% in Dice and +5.89% in IoU under identical re-implemented training conditions. The model retains 97.82% of its original performance across all nine perturbation conditions (robustness score: 0.6391), representing a 10.4% robustness improvement over the non-augmented baseline. These results demonstrate that the synergistic combination of attention mechanisms, transfer learning, and dual augmentation produces clinically viable robustness for facial wrinkle segmentation.