Automated inspection of objects undergoing general affine transformation

• US 6,421,458 B2
• Filed: 08/28/1998
• Issued: 07/16/2002
• Est. Priority Date: 08/28/1998
• Status: Expired due to Term

• Current Assignee: Cognex Corporation
• Original Assignee: Cognex Corporation
• Inventors: Michael, David J., Reyzin, Igor
• Primary Examiner(s): Boudreau, Leo
• Assistant Examiner(s): Werner, Brian P.
• Application Number:
  US09/141,932
• Publication Number:
  US 20010012395A1
• Time in Patent Office: 1,418 Days
• Field of Search: 382/151, 382/147-149, 382/216, 382/289, 382/294, 382/295, 382/296, 382/298, 348/87, 348/126, 702/40, 702/150, 702/151, 702/159, 250/559.3, 250/559.34, 250/559.37, 250/559.39, 250/559.46, 356/375, 356/237.4, 356/237.5
• US Class Current: 382/151
• CPC Class Codes: G06F 18/00 Pattern recognition
• CPC Class Codes: G06T 2207/30108 Industrial image inspection
• CPC Class Codes: G06T 7/001 using an image reference approach
• CPC Class Codes: G06T 7/33 using feature-based methods

Abstract

During statistical training and automated inspection of objects by a machine vision system, a General Affine Transform is advantageously employed to improve system performance. During statistical training, the affine poses of a plurality of training images are determined with respect to an alignment model image. Following filtering to remove high frequency content, the training images and their corresponding affine poses are applied to an affine transformation. The resulting transformed images are accumulated to compute template and threshold images to be used for run-time inspection. During run-time inspection, the affine pose of the run-time image relative to the alignment model image is determined. Following filtering of the run-time image, the run-time image is affine transformed by its affine pose. The resulting transform image is compared with the template and threshold images computed during statistical training to determine object status. In this manner, automated training and inspection is relatively less demanding on system storage, and results in an improvement in system speed and accuracy.

Claims

1. A method for automated inspection of an object comprising:

imaging an object to generate a run-time image;

determining the affine pose of the run-time image with respect to an alignment model image, created from a plurality of training samples used to form a template, the alignment model image representing the object at a single predetermined orientation wherein said affine pose comprises parameters defining said object'"'"'s orientation with at least four degrees of freedom;

prefiltering the run-time image using a filter that reduces errors introduced by a subsequent affine transformation to generate a filtered image;

transforming the filtered image with the affine pose to generate a transformed image;

mean-correcting the transformed image with said template image to provide a mean-corrected image;

comparing the mean-corrected image with a threshold image to produce an error image; and

analyzing the error image to determine object status.

2. The method of claim 1 further comprising determining a template and threshold image for the object.

3. The method of claim 2 wherein the template image comprises an average image.

4. The method of claim 2 wherein the threshold image comprises a function of a standard deviation image.

5. The method of claim 1 wherein determining the affine pose comprises determining the affine transform parameters which map the run-time image to the alignment model image.

6. The method of claim 1 wherein prefiltering comprises convolving the run-time image with a kernel adapted to eliminate high-frequency elements from the image.

7. The method of claim 6 wherein the kernel comprises an impulse function.

8. The method of claim 1 wherein transforming comprises applying the filtered run-time image to a General Affine Transform, the parameters of which are determined by the affine pose.

9. The method of claim 1 wherein mean correcting and comparing comprise the technique of double subtraction of the run-time image on a pixel-by-pixel basis, the output of which is an error image represented by:

10. The method of claim 1 wherein the template image comprises an average image, and wherein the threshold image comprises a linear transformation of the standard deviation of the average image.

11. The method of claim 1 wherein the template and threshold images are determined by:

iteratively imaging training objects to generate a plurality of training images;

determining the affine pose of each training image with respect to an alignment model image;

prefiltering each training image to generate filtered training images;

transforming each of the filtered training images with the corresponding affine pose to generate a plurality of transformed training images; and

computing a template image and threshold image of the object from the plurality of transformed training images.

12. The method of claim 11 further comprising generating an alignment model image from a geometric model of the object.

13. The method of claim 11 further comprising generating an alignment model image by selecting a portion of one of the training images as an alignment model image.

14. The method of claim 11 wherein determining the training image affine pose comprises determining the affine transform parameters which map the training image to the alignment model image.

15. In an artificial vision system, a method for statistical training of the system on an object comprising:

iteratively imaging an object to generate a plurality of training images;

determining the affine pose of each training image with respect to an alignment model image selected from one of the plurality of training samples, wherein said affine pose comprises parameters defining said object'"'"'s orientation with at least four degrees of freedom;

prefiltering each training image using a filter that reduces errors introduced by a subsequent affine transformation to generate filtered images;

transforming each of the filtered images with the corresponding affine pose to generate a plurality of transformed images;

computing a template image and threshold image of the object from the plurality of transformed images; and

comparing the alignment model with a run-time image of the object to align the run-time image with the template.

16. The method of claim 15 further comprising generating an alignment model image from a geometric model of the object.

17. The method of claim 15 further comprising generating an alignment model image by selecting a portion of one of the training images as an alignment model image.

18. The method of claim 15 wherein determining the affine pose comprises determining the affine transform parameters which map the training image to the alignment model image.

19. The method of claim 15 wherein prefiltering comprises convolving the training images with a kernel adapted to eliminate high frequency elements from the image.

20. The method of claim 19 wherein the kernel comprises an impulse function.

21. The method of claim 15 wherein the template image and threshold image are each computed in a single accumulator.

22. The method of claim 15 wherein the template image comprises an average image.

23. The method of claim 15 wherein the threshold image comprises a linear function of a standard deviation image.

24. The method of claim 15 wherein transforming comprises applying the filtered run-time image to a General Affine Transform, the parameters of which are determined by the affine pose.

25. The method of claim 15 further comprising:

during run-time, imaging a run-time object to generate a run-time image;

determining the affine pose of the run-time image with respect to the alignment model image;

prefiltering the run-time image to generate a filtered run-time image;

transforming the filtered run-time image with the affine pose of the run-time image to generate a transformed run-time image;

mean-correcting the transformed run-time image with the template image to provide a mean-corrected image;

comparing the mean-corrected image with the threshold image to produce an error image; and

analyzing the error image to determine object status.

26. A system for automated inspection of an object comprising:

an imaging system for imaging an object to generate a run-time image;

an alignment unit for determining the affine pose of the run-time image with respect to an alignment model image created from a plurality of training samples used to form a template, the alignment model image representing the object at a single predetermined orientation wherein said affine pose comprises parameters defining the object'"'"'s orientation with at least four degrees of freedom;

a filter for prefiltering the run-time image to generate a filtered image using a filter that reduces errors introduced by subsequent affine transformation;

an affine transform for transforming the filtered image with the affine pose to generate a transformed image;

a mean-corrector for correcting the transformed image with said template image to provide a mean-corrected image;

a comparator for comparing the mean-corrected image with the a threshold image to produce an error image;

an analyzer for analyzing the error image to determine object status; and

means for determining a template and threshold image for the object.

27. The system of claim 26 wherein the affine pose comprises affine transform parameters which map the run-time image to the alignment model image.

28. The system of claim 26 wherein the mean-correcting and comparator perform double subtraction of the run-time image on a pixel-by-pixel basis, the output of which is an error image represented by:

29. The system of claim 26 wherein the template image comprises an average image, and wherein the threshold image comprises a linear transformation of the standard deviation of the average image.

30. The system of claim 26 further comprising a system for determining the template and threshold images comprising:

an imaging system for iteratively imaging training objects to generate a plurality of training images;

an alignment unit for determining the affine pose of each training image with respect to an alignment model image;

a training filter for prefiltering each training image to generate filtered training images;

a training affine transform for transforming each of the filtered training images with the corresponding affine pose to generate a plurality of transformed training images; and

means for computing a template image and threshold image of the object from the plurality of transformed training images.

31. The system of claim 30 wherein the alignment model image is generated from a geometric model of the object.

32. The system of claim 30 wherein the alignment model image is generated by selecting a portion of one of the training images as an alignment model image.

33. The system of claim 30 wherein the training image affine pose comprises the affine transform parameters which map the training image to the alignment model image.

34. In an artificial vision system, a system for statistical training of the system on an object comprising:

an imaging system for iteratively imaging an object to generate a plurality of training images;

an alignment unit for determining the affine pose of each training image with respect to an alignment model image selected from one of the plurality of training samples, wherein said affine pose comprises parameters defining said object'"'"'s orientation with at least four degrees of freedom;

a filter for prefiltering each training image using a filter that reduces errors introduced by a subsequent affine transformation to generate filtered images;

an affine transform for transforming each of the filtered images with the corresponding affine pose to generate a plurality of transformed images;

means for computing a template image and threshold image of the object from the plurality of transformed images; and

means for comparing the alignment model with a run-time image of the object to align the run-time image with the template.

35. The system of claim 34 wherein the alignment model image is generated from a geometric model of the object.

36. The system of claim 34 wherein the alignment model image is selected form one of the training images.

37. The system of claim 34 further comprising a template image accumulator and a threshold image accumulator for computing a single template representation of the object.

38. The system of claim 34 wherein, during run-time, the imaging system further images a run-time object to generate a run-time image, the system further comprising

a run-time alignment unit for determining the affine pose of the run-time image with respect to the alignment model image;

a run-time filter for prefiltering the run-time image to generate a filtered run-time image;

a run-time affine transform for transforming the filtered run-time image with the affine pose of the run-time image to generate a transformed run-time image;

a mean-corrector for correcting the transformed run-time image with the template image to provide a mean-corrected image;

a comparator for comparing the mean-corrected image with the threshold image to produce an error image;

an analyzer for analyzing the error image to determine object status.