Toward 0-Norm Reconstruction, and Nullspace Technique for Compressive Sampling
Christine Law
EE, Stanford University
Talk
Dr. Law obtained her PhD from Stanford University in 2009. She is an expert in non-standard (spiral) MRI sequence design, with applications to fMRI, and in nonlinear image reconstruction using sparsity-enforcing penalties. At present, she is a research assistant in the group of Prof. Glover, Lucas MRS Imaging Center, Stanford University.
Compressive sampling (compressed sensing) conventionally means 1-norm approximation to 0-norm minimization. Advantages and limitations of the 1-norm technique and alternative methods for computing 0-norm solution will be presented.
Two fast 0-norm algorithms are introduced for imaging with application to Magnetic Resonance Imaging (MRI). Live Matlab demos of image reconstruction from highly undersampled images demonstrate their efficiency. These 0-norm techniques require fewer measurements than 1-norm.
We also demonstrate signal separation and perfect reconstruction from a highly undersampled composite 1D signal that is sparse with respect to two distinct dictionaries.
Deconvolution of haemodynamic response directly from signal data in functional MRI (fMRI) will be presented. This technique bypasses the conventional calibration step.
When 1-norm method fails, we show how cardinality-constraint problems can be solved more reliably by a new method utilizing measurement matrix nullspace.