Reconstruction of Finite Rate of Innovation Signals with Model-Fitting Approach

• Published in: IEEE transactions on signal processing Vol. 63; no. 22; pp. 6024 - 6036
• Main Authors: Dogan, Zafer, Gilliam, Christopher, Blu, Thierry, Van De Ville, Dimitri
• Format: Journal Article
• Language: English
• Published: IEEE 15.11.2015
• Subjects: Algorithms; Annihilating Filter; Biological system modeling; Cadzow; Computational modeling; Cramér-Rao's lower bound (CRLB); Errors; Estimation; finite-rate-of-innovation; Innovation; iterative quadratic maximum likelihood (IQML); Kernel; Kumaresan-Tufts; Mathematical analysis; Mathematical models; matrix pencil; model fitting; Noise; Reconstruction; Reconstruction algorithms; sampling; structured total least squares (STLS); Technological innovation; total least squares (TLS); Training
• ISSN: 1053-587X; 1941-0476
• DOI: 10.1109/TSP.2015.2461513

Finite rate of innovation (FRI) is a recent framework for sampling and reconstruction of a large class of parametric signals that are characterized by finite number of innovations (parameters) per unit interval. In the absence of noise, exact recovery of FRI signals has been demonstrated. In the noisy scenario, there exist techniques to deal with non-ideal measurements. Yet, the accuracy and resiliency to noise and model mismatch are still challenging problems for real-world applications. We address the reconstruction of FRI signals, specifically a stream of Diracs, from few signal samples degraded by noise and we propose a new FRI reconstruction method that is based on a model-fitting approach related to the structured-TLS problem. The model-fitting method is based on minimizing the training error, that is, the error between the computed and the recovered moments (i.e., the FRI-samples of the signal), subject to an annihilation system. We present our framework for three different constraints of the annihilation system. Moreover, we propose a model order selection framework to determine the innovation rate of the signal; i.e., the number of Diracs by estimating the noise level through the training error curve. We compare the performance of the model-fitting approach with known FRI reconstruction algorithms and Cramér-Rao's lower bound (CRLB) to validate these contributions.