Design Sparse Features for Age Estimation

Jinli Suo1,3     Tianfu Wu1     Shiguang Shan3         Xilin Chen3         Wen Gao3
1Lotus Hill Institute           2University of California, Los Angeles
3Institute of Computing Technology, Graduate University of Chinese Academy of Sciences,China


1. Inspiring Observations

local


     Local facial features contribute largely to age perception.
     Hair color and style also influence age perception.

2. Our Solution
     Considering the large influences of local facial features and hair appearances on age perception, we propose to adopt a compositional face representation to decompose the human faces into semantic subregions and extract local features for automatic age estimation. Making good use of prior knowledge of age perception, we design informative features at different subregions, thus the feature dimension is largely reduced. The experimental results on two aging databases show that designing feature set for age estimation under the guidance of hierarchical face model is a promising method and a flexible framework as well.

3. Hierarchical Face Decomposition
     In the adopted hierarchical graphical face model, and facial details are represented at low, middle and high resolution from coarse to fine. The building of hierarchical face model is guided by the empirical studies in anatomy and the fine arts.

facemodel

4. Feature Design on Prior Knowledge
    
In our approach, we extract four type of features, including:
  • Topologic features:
    •    describing the subclassed of each facial part
  • Geometry features:
    •    describing the geometric attributes of each facial part
  • Photometry features:
    •    describing the photometry of each facial part
  • Configuration features:
    •    describing the spatial relationship and other constraints among regions

    chain

    5. The Estimation Accuracy
        
  • Mean Abosolute Error(MAE):
    •    For frontal face images taken under normal enviorment, the mean absolute error is around 5 years;
  • Cummulative Score(CS):
    •    For frontal face images taken under normal enviorment, the correction rate with estimation error under 10 years is around 90%.

    6. Publications
         [1] Jinli Suo, Tianfu Wu, Song-Chun Zhu, Shiguang Shan, Xilin Chen and Wen Gao, "Design Sparse Features for Age Estimation Using Hierarchical Face Model", FG 2008. [PDF]