Chapter 13
Two-dimensional synthetic face
generation using set estimation
13.1 Introduction . . . . . . . . .. . . . . .. . . . .. . . . . .. . . . . .. . . . .. . . . . .. . . . .. . . . . . 273
13.2 Generating face points from intraclass face images . . . . . .. . . . . .. 274
13.2.1 Face generation using algorithm with intraclass features
and related peak signal to noise ratio . . . .. . . . . .. . . . .. . . . 274
13.3 Generating face points from interclass face images .. . .. . . . . .. . . 277
13.3.1 Face generation with interclass features . . .. . . . . .. . . . .. . 280
13.3.2 Rejection of the non-meaningful face and corresponding
PSNR test . . . . . . . . .. . . . . .. . . . . .. . . . .. . . . . .. . . . . .. . . . .. . . . 283
13.4 Generalization capability of set estimation method . .. . . . .. . . . . . 283
13.5 Test of significance . . . . . . . .. . . . .. . . . . .. . . . . .. . . . .. . . . . .. . . . . .. . . . . 285
13.1 Introduction
Synthetic generation of face images for the purpose of face recognition has
been explored in recent times. Two dimensional (2D) to three-dimensional
(3D) reconstruction and generation of new face images of various shapes and
appearances had received attention. The popular solution to the problem is
proposed by 2D and 3D modeling of faces. 3D models include face mesh
frames, morphable models and depth map-based models, where one needs to
incorporate high quality graphics and complex animation algorithms. Flynn
et al. [85] provided a survey of approaches and challenges in 3D, multi-
modal 3D and 2D face recognition. 3D head poses are derived from 2D to 3D
feature correspondences [86]. Face recognition based on fitting a 3D morphable
model is also proposed with statistical texture [87]. Some significant works
are discussed in Table 2.5. Four main approaches for 2D modeling are
active appearance models (AAMs) [80], manifolds [210], geometry-driven face
synthesis methods [68] including face animation[82] and expression mapping
techniques [108],[84].
2D to 2D face reconstruction was initially developed by Cootes et al.
[80] and their active appearance, generated from 2D face images, was one
of the powerful methods. Multi-view face reconstruction in 2D space is done
by manifold analysis. Geometry-driven face synthesis [108] and expression
273

Get Face Detection and Recognition now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.