A Quality-Aware and Multi-Scale Representation Network for Low-Quality Face Alignment

Abstract

Low-quality face images often suffer from blur, occlusion, large pose variations and illumination changes, which make facial landmark localization challenging. Existing HRNetbased face alignment methods preserve high-resolution features, but their fixed multi-scale fusion strategy is insufficient for samples with different degradation levels. Moreover, directly introducing structural guidance may be ineffective without reliable multi-scale representations. To address these issues, we propose QAMSR-Net, a quality-aware face alignment network based on HRNet-W18. Specifically, a Coordinated Multi-Scale Representation Module (C-MSRM) is designed to adaptively balance enhancement and correction responses for robust feature representation. On this basis, a Structure-aware Quality Guidance Module (SQGM) is introduced to generate component-level structural supervision from existing landmark annotations and constrain heatmap prediction. Experiments on 300W, COFW and WFLW show that QAMSR-Net consistently improves HRNet-W18, especially on blur, occlusion and large-pose subsets.