Phone: (+61 8) 6488 4775
Fax: (+61 8) 6488 1089
Sparse representation for hyper/multi-spectral image analysis
To capture an image, a conventional RGB camera grossly quantizes the scene's radiance which results in the loss of useful spectral information. Hyper/Multi-spectral imaging is able to preserve this information very precisely, which makes this emerging imaging modality an attractive choice in applications ranging from remote sensing to medical imaging.
In my thesis, I propose novel methods to accomplish three fundamental tasks in hyper/multi-spectral image analysis, namely: (a) super-resolution of the images, (b) un-mixing of the spectra and (c) classification of the spectra/image features. The proposed methods are theoretically grounded in convex, non-convex and Bayesian sparse regression frameworks. The developed techniques have been published in the top-ranked conferences and journals in the fields of Computer Vision and Remote Sensing. These venues include IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Geosciences and Remote Sensing, IEEE Conference on Computer Vision and Pattern Recognition and European Conference on Computer Vision.
Whereas hyper/multi-spectral cameras are able to preserve the spectral details precisely, as compared to the contemporary RGB cameras, their spatial resolution is very low due to hardware constraints. Currently, this is the bottleneck in the ubiquitous use of the hyper/multi-spectral imaging technology. The proposed methods of super-resolution remove this bottleneck by fusing the spatial information of the RGB cameras with hyper/multi-spectral images.
It is a natural phenomenon that the reflectance spectra of different materials on the Earth's surface get mixed together before they are sensed by an air/space-borne camera. An accurate un-mixing of the spectra is a vital prerequisite for effective analysis of hyperspectral images in remote sensing. The un-mixing methods proposed in the thesis achieved state-of-the-art performance on the real remote sensing images acquired by the NASA's remote sensing instruments.
Face, object and human-action recognition are long standing problems in Computer Vision. The thesis proposes sparse representation based techniques that can be generically applied to all of these problems. Moreover, these methods have also been enhanced for hyper/multi-spectral classification tasks.