Abstract
Traditional wavelet de-noising algorithms often struggle with non-stationary signals, frequently destroying critical detail components. This paper proposes an adaptive wavelet de-noising algorithm that establishes a simplified mathematical model for thresholding. By introducing an adaptive threshold function that adjusts based on the relationship between wavelet decomposition scales of signal and noise, the algorithm preserves original signal details more effectively. Parameters are optimized using an iterative algorithm. Simulation results demonstrate that this adaptive approach outperforms traditional methods in terms of Signal-to-Noise Ratio (SNR) and Mean Square Error (MSE).
Introduction
In signal processing, removing noise while preserving essential signal characteristics is a fundamental challenge. Wavelet de-noising is preferred for non-stationary signals due to its multi-resolution analysis capabilities. However, standard techniques like hard and soft thresholding possess inherent flaws. Hard thresholding often introduces discontinuities or artifacts, while soft thresholding tends to shrink coefficients excessively, leading to a constant bias that blurs signal peaks and fine details.
This paper addresses these limitations by proposing a "best of both worlds" adaptive threshold function that dynamically adjusts its behavior between hard and soft states depending on the noise characteristics at different decomposition scales.
Methodology
The proposed framework relies on three core components: a flexible threshold function, scale-dependent adaptivity, and optimal threshold selection.
Adaptive Threshold Function
The authors define a variable threshold function $\eta(x, th, m)$, where $m$ serves as a control parameter. As shown in the figure below, when $m=1$, the function approximates soft thresholding. As $m$ increases, the function converges toward a hard thresholding response.
Scale-Dependent Adaptivity
Rather than applying a uniform thresholding strategy, the algorithm calculates the parameter $m$ for each wavelet decomposition scale $j$. This calculation is based on the ratio of noise energy to total signal energy at that specific scale. At finer scales, where noise typically dominates, the algorithm leans toward hard thresholding to aggressively suppress noise. At wider scales, it shifts toward soft thresholding to preserve the underlying signal structure.
Optimization via Newton’s Method
To determine the optimal threshold ($th$), the algorithm utilizes Stein’s Unbiased Risk Estimate (SURE) to estimate the Mean Square Error (MSE). Newton’s Method is then applied to find the threshold that minimizes this error. This iterative numerical technique was chosen for its second-order convergence, providing faster results than standard gradient descent methods.
Experimental Results
The algorithm was evaluated using standard benchmarks, including "Heavy Sine" and "Bumps" signals, which test frequency de-noising and peak preservation respectively.
Key performance metrics confirmed the superiority of the adaptive method:
- SNR Improvement: On the Heavy Sine signal, the proposed method achieved an SNR of approximately 21.97 dB, significantly higher than the 17.13 dB achieved by hard thresholding.
- MSE Reduction: The mean square error was reduced to 0.0495, compared to 0.1927 for hard thresholding.
- Local Peak Preservation: The Local Peak Error (LPE) was only 1.96%, whereas hard thresholding resulted in nearly 5% error, validating the method's ability to maintain critical signal "mutations."
Conclusion
The proposed adaptive wavelet de-noising algorithm successfully balances noise suppression with detail preservation. By linking the threshold's hardness to the energy distribution across scales and optimizing via Newton's method, it provides a robust solution for processing non-stationary signals. Future work will focus on extending this model to non-Gaussian noise environments and exploring the impact of different wavelet basis functions on overall performance.
Summary updated