Welcome to My Homepage

 

Home

Research

CV  

Photos

Links

Research

 

Curvelet

Curvelets form a 1-tight frame which is near optimal in representing curved singularities. Each element is called a curvelet. Curvelets have different scales, orientations and locations. Unlike wavelets, which are tensor product of 1D wavelets, curvelets are intrinsically defined in 2D (and 3D). Most importantly, curvelets are needle-shaped atoms with a lot of orientations. Thus, 2D curvelets are efficient in representing curved singularities and 3D curvelets are efficient in capturing surface singularities. As we know that edges are the main characteristics for an image, curvelets thus have applications in image denoising, compression and quality assessment, etc.

Another nice property of curvelets is that they remain coherent waveforms under wave equation in a smooth medium. There is also research going on in related area, such as wave computations, seismic imaging, etc.

Reference:

Curvelet.org

H. Douma, Maarten V. de Hoop, `Leading-order seismic imaging using curvelets', Geophysics (2007) in print.

J.-L. Starck, M.K. Nguyen and F. Murtagh, `Wavelets and Curvelets for Image Deconvolution: a Combined Approach', 2003.

 

 

Sparse Approximation

When solving an underdetermined system of equations, there are infinite many solutions if no criterion is applied. Only by adding some kind of constraint, one could possible narrow down the solution to a unique one. Sparse constraint is found to benefit so many modern applications in signal and image processing. Sparse approximation is the problem of approximating an input signal by a linear combination of atoms of an over-complete dictionary with one kind of sparsity constraint. Sparsity is usually measured by ell-0 norm, the number of non-zero items, which results in NP hard optimization problems. As a result, relaxed forms of minimizations involving ell-1 or ell-p (p<1) norm come into a play. Traditional methods such as interior point methods are accurate but slow. Greedy methods are fast but may fail badly. Facing large scale problems, researchers are looking for new algorithms which are efficient and also accurate. On the other hand, we would also like to know under what conditions, the solution is unique. Both theoretical and numerical aspects are under investigation.

Reference:

Afred M. Bruckstein, David L. Donoho and Michael Elad, `From Sparse Solutions of Systems of Equations to Sparse Modeling of Signals and Images', SIAM Review, 2009.

 

 

Other

Image Quality Assessment (IQA) is the method of quantifying the quality of natural images. In most cases, images are distorted by one or several filters. IQA has applications in data management, comparison among digital camera lens, evaluation of image processing algorithms such as denoising, deblurring algorithms, etc. Since noise, blur affect the edges more than smooth parts in an image, we expect to design new algorithms based on curvelet transforms.

Reference:

Ji Shen, Qin Li, Gordon Erlebacher, `Curvelet based No-Reference Objective Image Quality Assessment', PCS, Chicago, 2009

 

 

Currently, I am doing research on using curvelet as a sub-dictionary for separating 2D and 3D data into parts carrying different kinds of information. The algorithms I try to apply on are from the family of shrinkage algorithms. The problems are how efficient and stable the algorithms are, what the criterions are for good separation and under what condition, we can perfectly separate the information as we desire.

Furthermore, I am interested in IRLS-based algorithm for sparse approximation. There are variations on the traditional IRLS algorithm in order to suit for sparse approximation problems. However, in spite of the fact that the methods were shown to be accurate in several applications numerically, not many theoretical results have been established for this kind of methods. Inspired by a recent paper on the IRLS for sparse recovery, I try to prove the convergence of an IRLS method for error-constrained sparse approximation.

The other paper our group is working on is the curvlet-based no reference method for IQA. The log-pdf of curvelet coefficients for fine scales forms Generalized Gaussian Distribution for natural images. This statistic fact leads to our IQA method based on curvelet.