Method for estimating parameters of a graph spectral filter using training data

Patent metadata

Field	Value
Publication number	US-9619755-B2
Application number	US-201314060894-A
Country	US
Kind code	B2
Filing date	Oct 23, 2013
Priority date	Oct 23, 2013
Publication date	Apr 11, 2017
Grant date	Apr 11, 2017

Abstract

A method processes a signal represented as a graph by first determining a graph spectral transform based on the graph. In a spectral domain, parameters of a graph filter are estimated using a training data set of unenhanced and corresponding enhanced signals. The graph filter is derived based on the graph spectral transform and the estimated graph filter parameters. Then, the signal is processed using the graph filter to produce an output signal. The processing can enhance signals such as images by denoising or interpolating missing samples.

First claim

We claim:

1. A method for processing an input signal represented as a graph, comprising:
employing a processor executing computer executable instructions stored on a computer readable memory to facilitate performing the steps of:
acquiring the input signal, wherein the input signal is an image that is an unenhanced image including noise;
acquiring a training data set, wherein the training data set includes pairs of training signals, wherein each pair of training signals includes an unenhanced training signal with noise and a corresponding enhanced training signal with reduced noise, and wherein the training signals are in a form of images;
determining a graph spectral transform as an orthogonal set of eigenvectors based on the graph, wherein the graph includes vertices connected by edges having associated weights, and wherein the weights are based on a guidance signal in a form of a guidance texture image;
estimating, in a graph spectral domain, parameters of a graph filter based on the training data set;
deriving the graph filter based on the graph spectral transform and the graph filter parameters; and
processing the input image using the graph filter to produce an output image, wherein the output image is an enhanced image with reduced noise that provides an increased understanding of a displayed image over an unenhanced displayed image with noise.

2. The method of claim 1 , wherein the estimating uses a least square procedure.

3. The method of claim 2 , wherein the least square procedure comprises:
obtaining an error image E from the pairs of training signals;
generating a image y based on the graph spectral transform U t and the error image E;
generating a matrix A from the training signals and the graph spectral transform U t ; and
estimating H 1 A † y, where † represents a pseudo-inverse of the matrix A.

4. The method of claim 3 , wherein the error image E is a difference between the enhanced image and the corresponding unenhanced images.

5. The method of claim 3 , wherein the error image E is the enhanced image.

6. The method of claim 1 , wherein the graph filter is H or UH 1 U t , where U t is the graph spectral transform.

7. The method of claim 1 , wherein the weights are joint bilateral weights.

8. The method of claim 1 , wherein the signal is a depth image, the guidance signal is a corresponding texture image of samples, and the weights are derived from a spatial distance between two samples and an intensity difference of two collocated samples in the corresponding texture image.

9. The method of claim 1 , wherein processing removes noise from the signal, and further comprising:
setting a scaling parameter ρ to a predetermined value ρ 0 ; and
constructing the graph filter as (I+ρH t H) −1 , wherein I is an identity matrix.

10. The method of claim 1 , wherein the processing interpolates the signal, and further comprises:
setting a scaling parameter ρ to a predetermined value ρ 0 ; and
constructing the graph filter as (J t J+ρH t H) −1 , where
J (m×N) =( I (M×N) |0 (M×(N−M) ),

where I is an identity matrix, M represents a number of known samples, and N a number of all samples.

11. The method of claim 1 , wherein the processing interpolates missing samples in the signal, and further comprising:
constructing the graph for each missing sample;
estimating the graph filter based on the graph for each missing sample to produce an estimated graph filter;
deriving the graph filter using the estimated graph filter to produce a derived graph filter; and
interpolating the missing sample using the derived graph filter.

12. The method of claim 11 , wherein the constructing of the graph comprises:
selecting known samples near to the missing sample; and
linking the missing sample to all known samples.

13. The method of claim 1 , wherein the processing interpolates missing samples in the signal, and further comprising:
partitioning the signal into non-overlapping patches;
constructing the graph for each patch;
estimating, for each patch, the graph filter based on the graph for the patch to produce an estimated graph filter;
deriving the graph filter using the estimated graph filter for each patch to produce a derived graph filter; and
interpolating the missing samples using the derived graph filters.

14. The method of claim 13 , wherein the graph is constructed by linking each missing sample to all known samples within the patch.

15. The method of claim 14 , wherein the graph has links between known samples.

16. The method of claim 1 , further comprising:
defining categories of the signal; and
estimating, for each category, the filter parameters H 1 of the graph filter to classify the signal as one of the categories.

17. The method of claim 16 , wherein the signal is an image including blocks, and classifying each block by detecting whether the block includes edges or a flat texture.

Assignees

• Mitsubishi Electric Res Laboratories Inc

Inventors

• Wang Yongzhe
• Tian Dong
• Mansour Hassan
• Vetro Anthony
• Ortega Antonio

Classifications

• G06N20/00 Primary

  Machine learning · CPC title

• G06V10/426

  Graphical representations · CPC title

• G06T2207/20072

  Graph-based image processing · CPC title

• G06K9/469

  Physics · mapped topic

• G06T2207/20081

  Training; Learning · CPC title