Loading...

Feature Extraction in Subspace Domain for Face Recognition

Safayani, Mehran | 2011

1200 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 42156 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Manzuri Shalmani, Mohammad Taghi
  7. Abstract:
  8. Feature extraction in subspace domain for face recognition has attracted growing attention in recent years. Face image shown by a long vector usually belongs to a manifold of intrinsically low dimension. Researchers in face recognition field try to extract these manifolds using algebraic and statistical tools. Recently, the use of multilinear algebra and multidimensional data in various stages of feature extraction and recognition is considered. This approach reduces small sample size problem and computational cost by considering the spatial information in the image. Although these successes, the performance of the methods based of this idea in term of recognition rate in the applications such as face recognition is not satisfying which is due to the how to representing the face images and also some model based problems such as heteroscedasticity. Moreover, the proper probabilistic model should be introduced for further extending of these approaches. In this thesis, we propose a new feature extraction technique called three dimensional modular discriminant analysis (3DMDA) which utilizes a new data representation model. In this model, each image is partitioned into the several equal size local blocks, and the local blocks are combined to represent the image as a third-order tensor. This model can better use of the spatial redundancies in the image. Then, a new optimization algorithm which discards the subspaces with no useful information through optimization phase is introduced. To address heteroscedastic problem which is due to the implicitly assumptions of the mutlilinear discriminant analysis (MDA) approach a method called heteroscedastic multilinear discriminant analysis (HMDA) is proposed. This method removes the assumption of equal covariance matrix for different columns of the face image. Also it is shown that this approach is a general form of the multilinear discriminant analysis method. Finally, a matrix variate probabilistic model for two-dimensional canonical correlation analysis has been proposed and its relationships with two-dimensional discriminant analysis has been investigated. This model can be used for extending the two-dimensional feature extraction techniques. The experimental results has been evaluated on the synthetic data and face databases such as AT& T, CMU-PIE, AR ,Yale, Jaffe and Sheffield in terms of classification accuracy and computational complexity and has been compared by state of the art approaches in this area. The results show that 3DMDA increases the recognition rate of the two-dimensional discriminant analysis up to 28.71%, the kernel methods up to 15.12% and Null-space LDA up to 10.87% . Also, it can be observed that the HMDA approach improves the recognition results of MDA up to 18.54%. These results demonstrate the efficiency of the proposed methods and their superiority over existing methods.
  9. Keywords:
  10. Feature Extraction ; Subspaces ; Dimension Reduction ; Face Recognition ; Heteroscedasticity ; Probabilistic Feature Extraction

 Digital Object List

 Bookmark

...see more