Michaël Unser: Signal processing and bioimaging

A functional framework for sparse, non-gaussian signal processing and bioimaging

In recent years, the research focus in signal processing has shifted away from the classical linear paradigm, which is intimately linked with the theory of stationary Gaussian processes. Instead of considering Fourier transforms and performing quadratic optimization, researchers are presently favoring wavelet-like representations and have adopted 'sparsity' as design paradigm. Our ambition is to develop a unifying operator-based framework for signal processing that would provide the ''sparse" counterpart of the classical theory, which is currently missing. To that end, we shall specify and investigate sparse stochastic processes that are continuously-defined and ruled by differential equations, and construct corresponding wavelet-like sparsifying transforms. Our hope is to be able to rigorously connect non-quadratic regularization and sparsity-constrained optimization to well-defined continuous-domain statistical models. We also want to develop a novel Lie-group formalism for the design of steerable, signal-adapted wavelet transforms with improved invariance and sparsifying properties, both in 2-D and 3-D. We shall use these tools to define new reversible image representations in terms of singular points (contours and keypoints) and to develop novel algorithms for 3-D biomedical image analysis. In close collaboration with imaging scientists, we shall apply our framework to the development of new 3-D reconstruction algorithms for emerging bioimaging modalities such as fluorescence deconvolution microscopy, digital holography microscopy, X-ray phase-contrast microscopy, and advanced MRI.

Max ERC funding: 2.11 million Euros
Duration: 60 months
Host institution: EPFL
Project acronym: FUN-SP
Domain: Physical and Engineering Sciences