PhDs and Funding Opportunities
Data61, a Business Unit of CSIRO, offers a significant number of scholarships in a stimulating environment in which outstanding students can excel. Additionally, Data61 partners with universities on collaborative research projects which align with its strategic goals.
The Smart Vision Systems Group, which I am member of, is based in Canberra and is affiliated to the ANU and UNSW, two of Australia’s leading research universities. I am seeking high achieving and self-motivated students with an interest in the following areas to join our research team
Modeling Deep Learning Architectures
Deep learning architectures attempt to capture complex relationships in data by using multiple processing layers. Each of these layers are complex structures on their own right, each often comprised of non-linear transformations and linear operations on the inputs delivered by the previous layers. This project aims at obtaining and modelling generative relationships for these architectures. This is not only theoretically important, but practically useful since a better understanding of the complex relationships within and between layers can yield improvements in performance in applications such as recognition, regression, classification, etc. The project will focus in applications related to biometrics, such as aging analysis, expression recognition and gender and age recognition in big, real-world datasets.
Multispectral Pelican Imaging
Every scene comprises a rich tapestry of light sources, material reflectance, lighting effects due to object curvature and shadows. Despite being reasonably effective for scene analysis, trichromatic (i.e. RGB) technology does have limits in its scene analysis capabilities. In contrast, multispectral imaging delivers multiple channels that permit a better representation of the surface radiance. Moreover, recently architectures such as extended mosaic arrays, pelican imagers and light-field cameras have become popular in the community. We will explore the use of hyperspectral pelican imagers and high-performance cameras to enable a wide range of higher-level scene analysis features such as touch and focus, autofocusing, white balance and BRDF estimation for post and pre processing image tasks, digital content generation and computational photography.
Compression of Hyperspectral Images
We will explore the use of lossless compression of hyperspectral images to enable mobil devices to transmit and receive imagery using limited computational resources and low transmission data rates. We also aim at developing methods that allow basic hyperspectral image processing techniques to be applied to the compressed data and, if possible, apply existing standards and toolboxes so as to be used as a basis for further development. This is important since image processing on compressed images has shown to greatly reduce memory requirements while increasing computational efficiency by avoiding decompression and space domain operations.
Spectral Imaging in Scattering Media
Imaging spectroscopy has found growing applications in underwater, reduced visibility environments (fog, snow, smoke, etc.) and aerial imaging, where scattering effects should be taken into account in the imaging process. The information provided by multispectral imaging can be used to derive physical models for the appearances of a surface immersed in a scattering medium. This opens-up opportunities regarding image enhancement, navigation, underwater operations and low-light conditions.
Spectral Unmixing and Material Discovery
Hyperspectral unmixing is a crucial pre-processing step for material classification and recognition. This is as it involves estimating all or some of “end members” that make up the spectral signatures in a hyperspectral or multispectral image. This provides a means to material discovery using machine learning techniques or constrained optimisation methods to be used when end member dictionaries are available. As a result, many hyperspectral unmixing methods have been proposed in recent years. Of particular interest here are unmixing methods (either semi-supervised or unsupervised) which employ sparse regression and nonlinear machine learning approaches based upon statistics. We will also explore the use of manifold optimisation for hyperspectral unmixing.
If you are interested in any of the above topics, feel free to contact me. Both, Australian citizens/permanent residents and overseas applicants are encouraged to apply. The successful applicants will be expected to commence in March 2017 and fulfil all the ANU admission requirements. Applications close for international students on Wednesday 31 August 2016 and for Australian domestic students (Australian citizens/permanent residents) on Monday 31 October 2016.
