HP-FIBA: A Spatial Attention Model and Dataset for Facial Image Blurriness Assessment
Haoxin Luo, Jiandong Jin, Junyuan Yuan, Zhen Sun, Chang Kong, Xian Li · 2025
In this paper, we introduce HP-FIBA, a large-scale dataset specifically designed for perceptual facial blurriness assessment, comprising 115,055 images annotated through a multi-tier cyclic annotation (MTCA) method to ensure continuous and reliable blur labels. To leverage the perceptual characteristics of facial blurriness, we propose a spatial attention model that focuses on key facial regions while incorporating label distribution learning for more precise blur estimation. Extensive experiments demonstrate that region-based spatial attention significantly enhances assessment accuracy. Furthermore, cross-domain evaluation reveals a domain gap, emphasizing the need for adaptation strategies. Finally, evaluations on external datasets confirm the model's strong generalization performance.