Multi-class mixture of linear SVM methods for pattern recognition on Hyperspectral data

Pattaraporn Khuwuthyakorn (NICTA)

COMPUTER VISION AND ROBOTICS SERIES HDR Monitoring

DATE: 2011-05-05
TIME: 16:00:00 - 17:00:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
This presentation is for CECS HDR monitoring. Generally, it contains a brief of my project, progress, current status, achievements and plans as a progress report. The presentation will also detail about Multi-class mixture of linear SVM methods (multi-class MLSVM) extended from the binary classification version, which are apart of the 2nd phase of my project. Both binary and multi-class methods have been modelled by combining mixtures of linear support vector machine to get new classification methods. Basically, they divide features into subregions; each can be linearly separable by a linear SVM partioning. By viewing the feature space as a generation of the multivariate Gaussian distribution and establishing a link to the SVM-error related probabilities with the joint probabilistic distribution, MLSVM has a probabilistic interpretion, which allows learning parameters of the model using the EM optimization. The result on synthetic problem and on real-world hyperspectral data will also be illustrated.
BIO:
Pattaraporn Khuwuthyakorn is a PhD student at the ANU. Her research interests are in the areas of computer vision and pattern recognition. She is working on the NICTA Smart Trap project, which is a joint project between the Cooperative Research Centre for National Plant Biosecurity (CRCNPB) and NICTA. It aims at developing algorithms to detect and classify plant pests based on hyperspectral imaging.