Structural guidance enhances attention networks for multi-dimensional facial skin quality assessment

Accurate assessment of facial skin condition is essential for both dermatological diagnostics and the development of personalized skincare solutions. However, existing image-based methods are insufficiently sensitive to fine-scale variations and are easily affected by illumination. We propose SGA-U-Net, which couples structural guidance (SG) with structure-aware attention (SGA). Rule-based pseudo-labels of pores and wrinkles derived via classical image processing provide auxiliary supervision, and attention-augmented skip connections jointly predict four skin quality indicators, moisture, glossiness, sebum, and elasticity within a multi-task framework. Validated on our high-resolution, dual-illumination FaceSkin-Q dataset and the public Korean Skin Condition Measurement Dataset (NIA), the method consistently surpasses convolutional neural network (CNN) and Transformer baselines in both accuracy and illumination robustness. These results indicate that combining structural priors with attention enhances sensitivity to skin microstructures and improves cross-scene stability, suggesting potential utility for supporting future dermatological assessment and personalized skincare applications. Code is publicly available at https://github.com/CISLAB-web/SGAUnet .