BibTeX entry

@PHDTHESIS{201009Shervin_Rahimzadeh_Arashloo,
  AUTHOR={Shervin Rahimzadeh Arashloo},
  TITLE={Pose-Invariant 2D Face Recognition by Matching Using Graphical Models},
  SCHOOL={University of Surrey},
  MONTH=Sep,
  YEAR=2010,
  URL={http://www.bmva.org/theses/2010/2010-rahimzadeharashloo.pdf},
}

Abstract

The thesis presents a 2D face recognition system using Markov random field matching methodology for establishing dense correspondences between a pair of images in the presence of pose changes and self-occlusion. The proposed method, which exploits both shape and texture differences between images, achieves very competitive performance compared to the current approaches. The algorithm bypasses the need for geometric pre-processing of face images. By virtue of the matching methodology embedded in the algorithm, the proposed approach can cope with moderate translation, in and out of plane rotation, scaling and perspective effects. Also by employing a graphical model based approach, the proposed system circumvents the need for non-frontal images being available for training a pose-invariant face recognition system. In contrast to the state-of-the-art approaches based on 3D models, the approach operates on 2D images and bypasses the need for 3D face training data and avoids the vagaries of 3D face model to 2D face image fitting. From the point of view of object recognition based on graphical models, the matching energy in graph based approaches is shown to exhibit certain drawbacks and should not be used as a similarity criterion for the hypothesis selection directly. The main shortcomings of the energy functional (using at most pairwise potentials) are identified and a plausible energy normalization scheme is proposed. In order to reduce the computational burden of the inference in the model, two multi-scale processing approaches are proposed. One employs the super-coupling transform in order to solve the matching problem in a multiresolution fashion. The other is heuristic but surprisingly leads to good results. Last but not least, a sparse graphical model for facial feature localization is proposed. The method takes advantage of the sparsity of facial image features in order to speed-up the matching process. The conditional dependencies between different groups of image primitives are included as higher order interactions based on point distribution models and linearity-based priors. The sparse model has been successfully applied to the task of facial feature localization and also as an initialization step to speed-up inference in a more costly matching approach.