Signal Processing and Communications Laboratory

Department of Engineering

Timothy RobertsIMG_8611-HDR.jpg

Background - Research - Papers - Talks - Teaching

Position: PhD Student

E-mail: tr331 [at]

Office Location: BN3-04

Thesis Title: Sparsity-Based Methods in Image Processing

Supervisor: Nick Kingsbury

Advisor: Joan Lasenby


I have a BS in Electrical and Computer Engineering (2007) and an MEng in Information Engineering (2009), both from Cornell University. Between degrees I've worked in quantitative finance and as an embedded software engineer. Taking these experiences in statistics and digital signal processing together, I converged on an exciting academic adventure at Cambridge which combines my interests.

Research Interests

Sparse recovery, probabilistic modeling, and classification in image processing. Time series analysis. Bayesian methods.


Working Papers/To Appear

link to pdf Image Deconvolution Using Tree-Structured Bayesian Group Sparse Modeling

Using a Variational approximation, we derive a group-sparsity inducing algorithm which is accelerated using Majorization Minimization and show superior performance of parent-children and parent-child grouping strategies to a singleton grouping strategy in the context of deconvolution.

Ganchi Zhang, Timothy Roberts, Nick Kingsbury
IEEE International Conference on Imaging Processing, 2014


link to pdf Fast Approximate L0-norm Deconvolution Using Structured Wavelet Domain Priors

Regularizing the solution space of ill-posed problems using sparsity in the wavelet domain can be improved by sparsifying at the group level, which has the advantage of promoting another important property: persistence across scales. Here we show how this can be done in an L0-sense by using a re-weighted L2 penalty on groups of overlapping coefficients. The variable duplication approach leads to a locally-convex penalty by following a Bayesian argument.

Timothy Roberts, Nick Kingsbury
IEEE International Conference on Acoustics, Speech, and Signal Processing, 2014
pdf | poster

link to pdf Sparse Recovery of Phase-Encoded Velocity Images using Iterative Thresholding

In this paper we derive a new algorithm for improved reconstruction of phase-encoded velocity images obtained from compressively-sampled NMR acquisitions. Such images are used to characterize and improve the design and yield predictions from fixed-bed catalytic chemical reactors, which are widely used in petrochemical applications. Regularizing the real and imaginary parts of the acquired complex image separately using complex-wavelet transform-domain sparsity produces a piecewise-smooth velocity map, which is a substantial improvement over prior total-variation based methods.

Timothy Roberts, Nick Kingsbury, Daniel J Holland
IEEE International Conference on Image Processing, 2013
pdf | poster


link to pdf Bayesian Denoising/Deblurring of Poisson-Gaussian Corrupted Data Using Complex Wavelets

As part of the ISBI '14 Deconvolution Challenge we use a non-stationary variance, non-stationary mean prior to deconvolve 3D fluorescence microscope volumes which have been blurred and corrupted by two independent noise sources (Poisson and Gaussian). Our MAP Bayesian estimator is efficient to compute and more accurate than results obtained using the standard Richardson Lucy algorithm.

Timothy Roberts, Nick Kingsbury
IEEE International Symposium on Biomedical Imaging, 2014
pdf | poster | slides (embedded animations)


Supervised 3F1 - Signals and Systems (2013, 2014)

Demonstrated 3F4 - Data Transmission (2013, 2014)