Reconstruction of Fri Signals Using Deep Neural Network Approaches

Vincent C. H. Leung, Junjie Huang, Pier Luigi Dragotti

Abstract

Finite Rate of Innovation (FRI) theory considers sampling and reconstruction of classes of non-bandlimited continuous signals that have a small number of free parameters, such as a stream of Diracs. The task of reconstructing FRI signals from discrete samples is often transformed into a spectral estimation problem and solved using Prony's method and matrix pencil method which involve estimating signal subspaces. They achieve an optimal performance given by the Cramer-Rao bound yet break down at a certain peak signal-to-noise ratio (PSNR). This is probably due to the so-called subspace swap event. In this paper, we aim to alleviate the subspace swap problem and investigate alternative approaches including directly estimating FRI parameters using deep neural networks and utilising deep neural networks as denoisers to reduce the noise in the samples. Simulations show significant improvements on the breakdown PSNR over existing FRI methods, which still outperform learning-based approaches in medium to high PSNR regimes.

Introduction

Finite Rate of Innovation (FRI) theory provides a powerful framework for sampling and perfectly reconstructing non-bandlimited signals characterized by a finite number of degrees of freedom per unit of time. A quintessential example is a stream of Dirac pulses, which can be fully described by their locations and amplitudes. Traditionally, reconstructing these signals from discrete samples involves mapping the problem to a spectral estimation task, typically solved by subspace-based algorithms like Prony’s method or the Matrix Pencil method. While these methods are mathematically elegant and statistically efficient—reaching the theoretical Cramér-Rao Bound (CRB) in low-noise conditions—they exhibit a catastrophic failure known as "breakdown" when the signal-to-noise ratio (PSNR) falls below a specific threshold.

The Subspace Swap Problem

The breakdown in traditional FRI reconstruction is primarily attributed to "subspace swap" events. In these scenarios, the noise power becomes comparable to the signal power, causing the algorithm to incorrectly identify a noise-dominated dimension as part of the signal subspace. This geometric failure leads to wild, non-linear errors in estimating parameters like pulse locations ($t_k$).

Standard deviation of pulse location estimates versus PSNR, highlighting the breakdown point where traditional methods deviate from the Cramér-Rao Bound (CRB).

As illustrated in the performance curves, traditional methods like Matrix Pencil follow the CRB closely until they hit a critical PSNR, after which the error increases exponentially. This research aims to bypass this inherent limitation by leveraging the non-linear mapping capabilities of Deep Neural Networks (DNNs).

Deep Learning for FRI Reconstruction

To address the instability of subspace methods, the authors investigate two distinct deep learning strategies:

  1. Direct Inference (End-to-End): A DNN is trained to learn the direct mapping from noisy discrete samples to pulse locations. This approach is particularly advantageous because it implicitly learns the characteristics of the sampling kernel from the data, removing the need for explicit mathematical modeling of the acquisition hardware.
  2. DNN-based Denoising (Hybrid): In this configuration, the DNN acts as a pre-processor. It takes noisy samples and outputs "cleaned" versions, which are then fed into a traditional Matrix Pencil algorithm. This aims to combine the noise robustness of deep learning with the high precision of classical methods.

Architecture and Methodology

Both approaches utilize a similar backbone: three convolutional layers (100 filters each) followed by three fully connected layers. The models were trained on a synthetic dataset of $10^5$ pairs of two-Dirac signals ($K=2$) sampled with an exponential reproducing kernel (eMOMS). The signals were corrupted with additive white Gaussian noise across a wide range of PSNRs, from -5 dB to 70 dB.

Performance Analysis

Simulations reveal a significant trade-off between precision and robustness. Traditional subspace methods remain superior in high-PSNR (clean) regimes because they converge to the CRB. However, DNN approaches demonstrate far greater stability in high-noise environments.

Scatter plot of pulse location estimates ($t_k$) using the Direct Inference DNN approach across various PSNR levels, demonstrating stability in high-noise regimes.

In scenarios where pulses are very close together, the direct inference DNN significantly lowers the breakdown threshold. For instance, the DNN remains operational down to 30 dB, whereas the Matrix Pencil method fails at 45 dB—a massive 15 dB improvement in the functional range. While the DNN does not reach the extreme precision of classical methods in clean conditions due to generalization limits, its "graceful degradation" makes it a much more robust choice for practical, noisy applications like medical imaging or radar.

Conclusion

Deep learning offers a promising path to overcoming the geometric failures of traditional FRI reconstruction. By avoiding the subspace swap problem, DNNs provide stable estimates even under heavy noise. Future work will focus on improving high-PSNR precision through more complex architectures and extending these models to handle a variable or unknown number of pulses.

Summary updated

Abstract

Finite Rate of Innovation (FRI) theory considers sampling and reconstruction of classes of non-bandlimited continuous signals that have a small number of free parameters, such as a stream of Diracs. The task of reconstructing FRI signals from discrete samples is often transformed into a spectral estimation problem and solved using Prony's method and matrix pencil method which involve estimating signal subspaces. They achieve an optimal performance given by the Cramer-Rao bound yet break down at a certain peak signal-to-́ noise ratio (PSNR). This is probably due to the so-called subspace swap event. In this paper, we aim to alleviate the subspace swap problem and investigate alternative approaches including directly estimating FRI parameters using deep neural networks and utilising deep neural networks as denoisers to reduce the noise in the samples. Simulations show significant improvements on the breakdown PSNR over existing FRI methods, which still outperform learning-based approaches in medium to high PSNR regimes.
BibTeX
@inproceedings{leung2020reconstruction,
  author = {Vincent C. H. Leung and Junjie Huang and Pier Luigi Dragotti},
  title = {{Reconstruction of Fri Signals Using Deep Neural Network Approaches}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=4zCFXdHHdR7},
}

Loading related content...

Figure