Comprehensive Evaluation of Deep Learning Models for Facial Skin Segmentation

Skin segmentation is a crucial preprocessing step in facial image analysis. Conventional landmark-based or coarse segmentation methods often fail to extract skin regions accurately, especially near boundaries like the hairline. We comprehensively evaluate deep learning-based semantic segmentation approaches for distinguishing skin from non-skin areas, emphasizing exclusion of adjacent structures such as hair and eyebrows. Three factors are investigated: network architecture (four encoder-decoder models), model complexity (reduced channel widths), and illumination variation (simulated color temperatures). The best model achieved a skin-class IoU of 0.9633 and mIoU of 0.9656, and maintained accuracy under reduced complexity (mIoU>0.9600 at 25% channel width). Predictions also remained consistent across simulated color temperature shifts. Some predictions delineated boundaries more precisely than ground truth annotations. These results demonstrate that deep learning enables accurate and fine-grained facial skin segmentation, supporting its use as a reliable preprocessing step for automated skin analysis and potential applications in dermatological assessment and facial video-based physiological monitoring.