Method for image alignment under non-uniform illumination variations

• US 6,744,933 B2
• Filed: 01/18/2001
• Issued: 06/01/2004
• Est. Priority Date: 01/18/2001
• Status: Active Grant

• Current Assignee: Siemens Medical Solutions USA Incorporated (Siemens AG)
• Original Assignee: Siemens Corporate Research Incorporated (Siemens AG)
• Inventors: Fang, Ming, Lai, Shang-Hong
• Primary Examiner(s): Wu, Jingge
• Application Number:
  US09/764,586
• Publication Number:
  US 20020126915A1
• Time in Patent Office: 1,230 Days
• Field of Search: 382/294-297, 358/450
• US Class Current: 382/294
• CPC Class Codes: G06T 2207/30108 Industrial image inspection
• CPC Class Codes: G06T 7/001 using an image reference approach
• CPC Class Codes: G06T 7/37 using transform domain methods

Abstract

A method for matching images includes the step of providing a template image and an input image. A Laplacian-of-Gaussian filtered log (LOG-log) image function is computed with respect to the template image and the input image to obtain a Laplacian-of-Gaussian filtered template image and a Laplacian-of-Gaussian filtered input image, respectively. An energy function is minimized to determine estimated geometric transformation parameters and estimated photometric parameters for the input image with respect to the template image. The energy function is formed by weighting non-linear least squared differences of data constraints corresponding to locations of both the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image. The estimated geometric transformation parameters and the estimated photometric parameters are output for further processing. The method allows for image matching under non-uniform illumination variations.

Claims

1. A method for matching images, comprising the steps of:

providing a template image and an input image;

computing a Laplacian-of-Gaussian filtered log (LOG-log) image function with respect to the template image and the input image to obtain a Laplacian-of-Gaussian filtered template image and a Laplacian-of-Gaussian filtered input image, respectively;

minimizing an energy function formed by weighting non-linear least squared differences of data constraints corresponding to locations of both the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image to determine estimated geometric transformation parameters and estimated photometric parameters for the input image with respect to the template image; and

outputting the estimated geometric transformation parameters and the estimated photometric parameters for further processing.

2. The method according to claim 1, further comprising the steps of:

extracting wavelet features from an image gradient corresponding to the input image;

identifying, with respect to the wavelet features, a nearest-neighbor feature vector from a set of training data, the training data obtained by simulating a geometrical transformation on the template image with geometric parameters of the nearest-neighbor feature vector being uniformly sampled from a given search space; and

generating an initial guess for the estimated geometric transformation parameters, based upon the geometric parameters of the nearest-neighbor feature vector,

wherein the initial guess is utilized by the minimizing step to determine the estimated geometric transformation parameters.

3. The method according to claim 1, wherein said minimizing step comprises the steps of:

selecting pixel locations in the Laplacian-of-Gaussian filtered template image having a largest reliability measure; and

computing gradients and qualities for the selected pixel locations.

4. The method according to claim 1, wherein the geometric parameters comprise at least one of a translation vector, a rotation angle, and a size scaling factor.

5. The method according to claim 1, wherein said minimizing step further comprises the steps of:

calculating a Hessian matrix and a gradient vector of the energy function based on an initial guess of the geometric transformation parameters;

updating the initial guess based on the calculating of the Hessian matrix and the gradient vector of the energy function; and

iteratively recalculating the Hessian matrix and the gradient vector of the energy function until an updated guess is within an acceptable increment from a previous updated guess.

6. The method according to claim 1, wherein said computing step comprises the steps of:

applying a Gaussian filter to the template image and the input image to obtain a Gaussian filtered template image and a Gaussian filtered input image, respectively; and

applying a Laplacian operation to the Gaussian filtered template image and the Gaussian filtered input image to obtain the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image, respectively.

7. The method according to claim 6, wherein the Gaussian filtered template image and the Gaussian filtered input image have reduced noise with respect to the template image and the input image, respectively, and the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image have reduced non-uniform illumination with respect to the template image and the input image, respectively.

8. A method for matching images, comprising the steps of:

providing a template image and an input image;

computing a Laplacian-of-Gaussian filtered log (LOG-log) image function with respect to the template image and the input image to obtain a Laplacian-of-Gaussian filtered template image and a Laplacian-of-Gaussian filtered input image, respectively;

generating an initial guess for estimated geometric transformation parameters of the input image;

minimizing an energy function formed by weighting non-linear least squared differences of data constraints corresponding to locations of both the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image to determine the estimated geometric transformation parameters and estimated photometric parameters for the input image with respect to the template image; and

outputting the estimated geometric transformation parameters and the estimated photometric parameters for further processing,

wherein the initial guess is utilized by the minimizing step to determine the estimated geometric transformation parameters.

9. The method according to claim 8, further comprising the steps of:

extracting wavelet features from an image gradient corresponding to the input image;

identifying, with respect to the wavelet features, a nearest-neighbor feature vector from a set of training data, the training data obtained by simulating a geometrical transformation on the template image with geometric parameters of the nearest-neighbor feature vector being uniformly sampled from a given search space; and

generating the initial guess for the estimated geometric transformation parameters, based upon the geometric parameters of the nearest-neighbor feature vector.

10. The method according to claim 8, wherein said minimizing step comprises the steps of:

selecting pixel locations in the Laplacian-of-Gaussian filtered template image having a largest reliability measure; and

computing gradients and qualities for the selected pixel locations.

11. The method according to claim 8, wherein the geometric parameters comprise at least one of a translation vector, a rotation angle, and a size scaling factor.

12. The method according to claim 8, wherein said minimizing step further comprises the steps of:

calculating a Hessian matrix and a gradient vector of the energy function based on the initial guess of the geometric transformation parameters;

updating the initial guess based on the calculating of the Hessian matrix and the gradient vector of the energy function; and

iteratively recalculating the Hessian matrix and the gradient vector of the energy function until an updated guess is within an acceptable increment from a previous updated guess.

13. The method according to claim 8, wherein said computing step comprises the steps of:

applying a Gaussian filter to the template image and the input image to obtain a Gaussian filtered template image and a Gaussian filtered input image, respectively; and

applying a Laplacian operation to the Gaussian filtered template image and the Gaussian filtered input image to obtain the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image, respectively.

14. The method according to claim 13, wherein the Gaussian filtered template image and the Gaussian filtered input image have reduced noise with respect to the template image and the input image, respectively, and the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image have reduced non-uniform illumination with respect to the template image and the input image, respectively.

15. A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for matching images, said method steps comprising:

providing a template image and an input image;

computing a Laplacian-of-Gaussian filtered log (LOG-log) image function with respect to the template image and the input image to obtain a Laplacian-of-Gaussian filtered template image and a Laplacian-of-Gaussian filtered input image, respectively;

minimizing an energy function formed by weighting non-linear least squared differences of data constraints corresponding to locations of both the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image to determine estimated geometric transformation parameters and estimated photometric parameters for the input image with respect to the template image; and

outputting the estimated geometric transformation parameters and the estimated photometric parameters for further processing.

16. The program storage device according to claim 15, further comprising the steps of:

extracting wavelet features from an image gradient corresponding to the input image;

identifying, with respect to the wavelet features, a nearest-neighbor feature vector from a set of training data, the training data obtained by simulating a geometrical transformation on the template image with geometric parameters of the nearest-neighbor feature vector being uniformly sampled from a given search space; and

generating an initial guess for the estimated geometric transformation parameters, based upon the geometric parameters of the nearest-neighbor feature vector,

wherein the initial guess is utilized by the minimizing step to determine the estimated geometric transformation parameters.

17. The program storage device according to claim 15, wherein said minimizing step comprises the steps of:

selecting pixel locations in the Laplacian-of-Gaussian filtered template image having a largest reliability measure; and

computing gradients and qualities for the selected pixel locations.

18. The program storage device according to claim 15, wherein the geometric parameters comprise at least one of a translation vector, a rotation angle, and a size scaling factor.

19. The program storage device according to claim 15, wherein said minimizing step further comprises the steps of:

calculating a Hessian matrix and a gradient vector of the energy function based on an initial guess of the geometric transformation parameters;

updating the initial guess based on the calculating of the Hessian matrix and the gradient vector of the energy function; and

iteratively recalculating the Hessian matrix and the gradient vector of the energy function until an updated guess is within an acceptable increment from a previous updated guess.

20. The program storage device according to claim 15, wherein said computing step comprises the steps of:

applying a Gaussian filter to the template image and the input image to obtain a Gaussian filtered template image and a Gaussian filtered input image, respectively; and

applying a Laplacian operation to the Gaussian filtered template image and the Gaussian filtered input image to obtain the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image, respectively.

21. The method according to claim 20, wherein the Gaussian filtered template image and the Gaussian filtered input image have reduced noise with respect to the template image and the input image, respectively, and the Laplacian-of-Gaussian filtered template image and the Laplacian-of-Gaussian filtered input image have reduced non-uniform illumination with respect to the template image and the input image, respectively.