Adaptive wavelet thresholding for multichannel signal estimation

Ian C. Atkinson, Farzad Kamalabadi, Douglas L. Jones, N. MinhProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2003 · DOI

TL;DR

This paper enhances a wavelet-based method for estimating 2D multichannel signals by incorporating an overcomplete wavelet expansion and an adaptive thresholding technique. These improvements lead to significant gains in signal-to-noise ratio (SNR), outperforming existing methods and offering a more robust solution for applications like hyperspectral imaging.

What Problem Does This Paper Address?

Estimating signals from noisy observations is a fundamental problem in many fields, including communications, medical imaging, and remote sensing. A particularly challenging scenario is the estimation of multichannel signals, where multiple related signals are observed simultaneously (e.g., different spectral bands of an image). A naive approach of processing each channel independently ignores valuable inter-channel correlations. While optimal methods like the Wiener filter exist, they often require extensive statistical knowledge of the signal and noise, and can be computationally prohibitive. This paper aims to improve upon a recently proposed wavelet-based technique that approximates the optimal estimator without needing explicit statistical knowledge, by exploring enhancements that further boost estimation quality.

Methods

The core of the proposed method builds upon a wavelet-based estimation scheme that uses a Discrete Fourier Transform (DFT) for channel decorrelation and a 2D Discrete Wavelet Transform (DWT) for spatial decorrelation. This effectively approximates the optimal Karhunen-Loève (KL) transform without requiring prior statistical knowledge. Noise is then suppressed by applying a wavelet-domain threshold.

The paper introduces two key enhancements:

  1. Overcomplete Wavelet Expansion: Instead of a standard, critically-sampled DWT, the authors propose using an overcomplete wavelet expansion (specifically, the "Algorithme à Trous"). This approach is shift-invariant, meaning that shifting the input signal doesn't drastically change the output, which can lead to more consistent results. While it increases computational complexity, it offers improved estimation quality.
  2. BayesShrink Adaptive Wavelet Threshold: The standard method applies a single threshold to all wavelet coefficients. This paper incorporates the BayesShrink adaptive threshold, which adjusts the threshold for each subband based on its estimated noise variance and the distribution of its coefficients. This method assumes that the detail subbands follow a Generalized Gaussian (GG) distribution, which is often a reasonable assumption for image data.

Datasets

The primary dataset used for evaluation is the "Desert" multispectral dataset. This dataset consists of 48 channels, each with spatial dimensions of 281x251 pixels. For processing, these images were zero-padded to 512x512. Noisy observations were generated by adding additive zero-mean Gaussian noise with varying variances.

Key Insights and Ideas

  • Wavelets as KL Approximations: The paper leverages the insight that wavelets can effectively approximate the KL transform, which is crucial for decorrelating signals without explicit statistical knowledge.
  • Overcompleteness for Robustness: Using an overcomplete wavelet expansion improves the shift-invariance of the estimation process, leading to more stable and slightly better results, even though it increases computational cost.
  • Adaptive Thresholding for Precision: The BayesShrink adaptive thresholding significantly outperforms fixed thresholding by tailoring the noise suppression to the characteristics of each subband, especially when subband coefficients follow a Generalized Gaussian distribution.
  • Interchannel Correlation is Key: The effectiveness of the wavelet-based methods, particularly compared to single-channel processing (like wiener2), highlights the importance of exploiting interchannel correlations for improved multichannel signal estimation.

Key Results

  • Overcomplete Wavelet Expansion: Improved SNR by 1 to 1.5 dB compared to the standard critically-sampled DWT.
  • BayesShrink Adaptive Threshold: Achieved SNR gains of 3 dB or more compared to the standard soft thresholding, particularly at low to moderate noise levels. For instance, at $\sigma_n^2 = 500^2$, the BayesShrink estimator provided a 35% improvement in SNR gain over the soft estimator.
  • Overall Performance: The BayesShrink estimator consistently outperformed the wiener2 baseline and the standard soft wavelet estimator across various noise levels and channel counts. The Soft-Overcomplete estimator also showed significant improvements over the Soft estimator.
  • Noise Variance Impact: As noise variance increased, the maximum possible SNR gain also increased, and robust estimation schemes like BayesShrink maintained their advantage.
  • Channel Count Impact: The wavelet-based estimators showed improved performance as the number of channels increased, demonstrating their ability to leverage interchannel correlations.

Limitations and Future Work

  • Computational Cost: The overcomplete wavelet expansion significantly increases computational complexity, potentially limiting its practicality for very large datasets.
  • Generalized Gaussian Assumption: The BayesShrink method relies on the assumption that detail subbands follow a Generalized Gaussian distribution. While this holds for many natural signals, deviations could impact performance.
  • Future Work Suggestion: The authors suggest that combining an overcomplete wavelet expansion with an adaptive wavelet-domain threshold (like BayesShrink) could yield even superior results, especially in applications where estimation quality is paramount (e.g., medical imaging). This combination, despite its higher computational cost, might be justified by the potential for significantly improved estimates.

Summary updated

Abstract

In this paper, we illustrate how a recently proposed wavelet-based estimation scheme for 2-D multichannel signals can utilize an overcomplete wavelet expansion or the BayesShrink adaptive wavelet-domain threshold to improve estimation results. The existing technique approximates the optimal estimator using a DFT and an orthonormal 2-D DWT to efficiently decorrelate the signal in both channel and space, and a wavelet-domain threshold to suppress the noise. Although this technique typically yields signal-to-noise ratio (SNR) gains of over 12 dB, results can be improved 1 to 1.5 dB by replacing the critically-sampled wavelet expansion with an overcomplete wavelet expansion. In addition, provided that the detail subbands of the original signal channels each obey a generalized Gaussian distribution, average channel SNR gains can be improved 3 dB or more using the BayesShrink adaptive wavelet-domain threshold.
BibTeX
@inproceedings{atkinson2003adaptive,
  author = {Ian C. Atkinson and Farzad Kamalabadi and Douglas L. Jones and N. Minh},
  title = {{Adaptive wavelet thresholding for multichannel signal estimation}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2003},
  doi = {10.1117/12.506397},
}
Figure