Signal processing method and system for noise removal and signal extraction

• US 20050265629A1
• Filed: 05/31/2005
• Published: 12/01/2005
• Est. Priority Date: 05/28/2004
• Status: Active Grant

• Current Assignee: Lawrence Livermore National Security LLC (Government of the United States of America)
• Original Assignee: Regents of the University of California (University of California)
• Inventors: Fu, Chi Yung, Petrich, Loren
• Granted Patent: US 7,519,488 B2
• Time in Patent Office: Days
• US Class Current: 382/275
• CPC Class Codes: G06F 18/00 Pattern recognition
• CPC Class Codes: G06F 2218/06 by applying a scale-space analysis, e.g. using wavelet analysis

Abstract

A signal processing method and system combining smooth level wavelet pre-processing together with artificial neural networks all in the wavelet domain for signal denoising and extraction. Upon receiving a signal corrupted with noise, an n-level decomposition of the signal is performed using a discrete wavelet transform to produce a smooth component and a rough component for each decomposition level. The n^th level smooth component is then inputted into a corresponding neural network pre-trained to filter out noise in that component by pattern recognition in the wavelet domain. Additional rough components, beginning at the highest level, may also be retained and inputted into corresponding neural networks pre-trained to filter out noise in those components also by pattern recognition in the wavelet domain. In any case, an inverse discrete wavelet transform is performed on the combined output from all the neural networks to recover a clean signal back in the time domain.

Claims

1. A signal processing method comprising:

receiving a signal corrupted with noise;

performing an n-level decomposition of said signal using a discrete wavelet transform to produce a smooth component and a rough component for each decomposition level;

inputting the n^th level smooth component into a corresponding neural network pre-trained to filter out noise in said component by pattern recognition in the wavelet domain; and

performing an inverse discrete wavelet transform on output from the neural network to recover a clean signal.

2. The signal processing method of claim 1, further comprising:

inputting the rough components of pre-selected decomposition levels into separate corresponding neural networks each pre-trained to filter out noise in a corresponding one of the inputted rough components by pattern recognition in the wavelet domain; and

performing an inverse discrete wavelet transform on the outputs from all the neural networks to recover a clean signal.

3. The signal processing method of claim 2,

wherein the rough components are pre-selected beginning in order with the highest decomposition levels.

4. The signal processing method of claim 2,

wherein the rough components of all decomposition levels are inputted into separate corresponding neural networks.

5. The signal processing method of claim 1,

wherein said signal has a Gaussian shape.

6. The signal processing method of claim 5,

wherein said discrete wavelet transform is a Daubechies-2 wavelet transform.

7. The signal processing method of claim 6,

wherein a four-level decomposition is performed using said Daubechies-2 wavelet transform.

8. The signal processing method of claim 1,

wherein each of said neural networks is a projection neural network.

9. The signal processing method of claim 1,

further comprising removing a baseline prior to performing the n-level decomposition of said signal.

10. A signal processing method comprising:

receiving a signal corrupted with noise;

decomposing the signal into a plurality of signal components using a predetermined transform;

inputting each of the plurality of decomposed signal components into a corresponding neural network pre-trained to filter out noise in the signal component by pattern recognition in the transform domain; and

performing an inverse transform on output from the neural networks to recover a clean signal in the time domain.

11. The signal processing method of claim 10,

wherein the plurality of signal components are filtered to remove noise.

12. A system comprising:

a discrete wavelet transformer capable of iteratively decomposing a signal into a plurality of decomposition levels each having a smooth component and a rough component;

at least one neural network(s) corresponding in number to a set of pre-selected components comprising a highest-level smooth component and a predetermined number of rough components, each neural network operatively coupled to the discrete wavelet transformer to receive a corresponding one of the pre-selected components, and pre-trained to filter noise in the wavelet domain from a corresponding one of the pre-selected components via pattern recognition techniques; and

an inverse discrete wavelet transformer capable of recovering a clean signal in the time domain from the combined outputs of the plurality of neural networks.

13. The system of claim 12,

wherein each of plurality of neural networks has been trained using noiseless signal data or ideal data derived from physical principle and such trained neural network after sufficient training will generalize its capability to recover signal from unseen corrupted data.