| Abstract: |  This thesis deals with the problem of recognizing gender of a person from an image of his/her face. One of the major problems in the face classification comes from a large image variance caused by unknown pose of the recognized face. Existing approaches for gender recognition are based on feature classifiers trained from examples. The feature classifiers are themselves not invariant against pose transformations of the input face. There are two common strategies to make the classifier invariant. The first strategy is based on registering the
input images prior to their classification. The second strategy is based on generating synthetic training examples by applying all pose transformations which do not change the class membership. In this thesis we propose a new approach based on using a structural classifier which treats the unknown face pose as an additional hidden parameter. The proposed structural classifier performs face registration and face classification simultaneously in one step. We formulate learning of the structural classifier from annotated examples as a convex optimization problem. We experimentally compare the proposed structural classifier against one baseline approach and one state-of-the-art approach on a large corpus of faces. The experiments show that the proposed structural classifier outperforms the state-of-the-art method achieving relative improvement of 10% in the classification accuracy. We also propose an improvement of the Bundle Method for Regularized Risk Minimization which is an optimization algorithm suitable for solving large instances of the learning problem. Preliminary results demonstrate that the improved BMRM algorithm can significantly reduced the number of iterations of the original method. The last result of this thesis is an open source library implementing the proposed classifier and its learning algorithm.
|
|---|