Face authentication with pose adjustment using support vector machines with a Hausdorff-based kernel

Face authentication is a biometric classification method that verifies the identity of a user based an image of their face. Accuracy of the authentication is reduced when the pose of the training face images is different than the testing image. This dissertation describes two methodologies which can increase the face authentication accuracy, if the training and testing images poses are different. The first method uses cascading trilinear tensors which adjust the pose of 2D images in a 3D space. By being able to morph the images in a 3D space, the training and testing images can be normalized to have the same pose. Using support vector machines (SVM) as the classifier, the second method uses a Hausdorff-based kernel embedded in the SVM decision function. The Hausdorff-based kernel has been shown to improve accuracy in object recognition. Using these two methods, the face authentication accuracy is improved over methods which use classic SVM kernels or do not use pose adjustment.