Gender Perception and Gender Conversion

Jinli Suo1,3     Liang Lin1,2     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. Research Objective

     (i) Male and females are largely different from each other in skin, hairstyle, facial components, facial hair, et al., but which are most important?
     (ii) Can we perform digital image editing to change the gender belongingship and preserve the identity of given individuals? If yes, how to do it?

2. Our Work
     This work integrate studies of gender classification in computer vision with the image fusion techniques in computer graphics succesffuly for some specific applications.

facemodel


Firstly, we adopt a compositional graphical face representation to decompose a human face into semantic regions, and analyze the difference between two gender groups on different aspects, including structure of graph node, texture attributes of each node, and the spatial relationships among nodes, et al.
Secondly, we perform gender conversion on the graphical attributes highly related with gender perception to transform the gender attributes towards the opposite gender groups, while keeping as high similarity with the given image as possible.

3. Which Attibutes Are Crucial for Gender Classification?
    
  • Structure
    • The structure information is described by the shape of facial components and the profiles along the contour, we model them with graphs. There exist subtle differences between the structure of two genders, thus and graph matching is adopted to select a target template in the opposite gender group and a series of graph editing operations are performed to the given graph representation.

    examples

  • Texture
    • In the graphicl model texture information is modeled with PCA models, in which some princpal components are distinctive for gender classification and some are not. We select the ones curcial for gender perception and transform the parameter towards the opposite gender.

    examples

  • Configuration
    • The statics shows that there exist differences between face shapes of two opposite gender groups(see the following figure), here the difference between two average face shapes of two gender groups are superimposed onto the given shape for gender conversion.

    examples

    4. How to Generate Seamless Image Esiting Results?
         Poisson image editing[1] is most famous for its good editing result and easy implementation an improved version of which, proposed by Wen[2], proposed a fusion method to fuse a low spatial resolution multispectral image and a high spatial resolution panchromatic image together successfully. We adopted Wen's algorithm to fuse the local gender conversion intermediate results at different resolutions seamlessly.

    Poisson

    5. Examplar Fusion Results

    examples

    6. Publications
         Jinli Suo, Liang Lin, Shiguang Shan, Xilin Chen and Wen Gao, “High Resolution Face Fusion for Gender Conversion”, to appear in IEEE Trans. on Systems Man & Cybernetics, Part A, 2010.

    References
         [1] P. Perez, M. Gangnet, and A. Blake. Poisson image editing. ACM Trans. Graphics, vol. 22, no. 3, pp. 313–318, Jul. 2003.
    [2] J. Wen, Y. Li, and H. Gong. Remote sensing image fusion on gradient field. In Proc. Int’l. Conf. Pattern Recognition, pages 643–646, 2006.