Methods and apparatus for optically determining the acceptability of products

• US 5,046,111 A
• Filed: 02/09/1989
• Issued: 09/03/1991
• Est. Priority Date: 02/09/1989
• Status: Expired due to Term

• Current Assignee: Philip Morris USA Incorporated (Altria Group, Inc.)
• Original Assignee: Philip Morris USA Incorporated (Altria Group, Inc.)
• Inventors: Maher, Robert J., Cox, Kenneth A.
• Primary Examiner(s): Boudreau, Leo H.
• Assistant Examiner(s): NOT, DEFINED
• Application Number:
  US07/308,739
• Time in Patent Office: 936 Days
• Field of Search: 382/8, 382/39, 382/14, 382/15, 364/552
• US Class Current: 382/143
• CPC Class Codes: B07C 5/10 measured by light-responsive means sorting according to colour B07C5/342
• CPC Class Codes: G01N 21/88 Investigating the presence of flaws or contamination
• CPC Class Codes: G01N 21/8851 Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges G01N21/8806 and G01N21/93 - G01N21/95692 take precedence; optical measurement of dimensions G01B11/00; optical scanning G02B26/10; image transformation G06T3/00; computerised image enhancement G06T5/00; image processing per se for flaw detection G06T7/0002
• CPC Class Codes: G06F 18/21 Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
• CPC Class Codes: G06T 2207/30164 Workpiece; Machine component
• CPC Class Codes: G06T 7/001 using an image reference approach
• CPC Class Codes: G06V 10/70 using pattern recognition or machine learning optical pattern recognition or electronic computations therefor G06V10/88
• CPC Class Codes: G07C 3/14 Quality control systems

Abstract

The appearance of a product is determined to be acceptable or unacceptable by computing a discriminant function or image from a plurality of two-dimensional images of products of the kind to be inspected. A two-dimensional image of the product is then formed and processed using the discriminant image to produce an output value which can be compared to a predetermined value used in computing the discriminant image. If the output value deviates from the predetermined value by more than a predetermined amount, the product has been found to have an unacceptable appearance.

Claims

1. The method of determining whether or not a surface of an object has substantially the same appearance as the corresponding surface of a plurality of training objects, said method comprising the steps of:

for each training object, (a) forming a two-dimensional image of said surface, (b) subdividing said two-dimensional image into a plurality of pixels, (c) digitizing each of said pixels by associating with each pixel an initial digital value proportional to an optical characteristic of the associated pixel, (d) edge detecting said initial digital values by processing said initial digital values to produce output digital values such that initial digital values associated with pixels at or near substantial changes in initial digital value are relatively emphasized to produce the associated output digital values, while initial digital values associated with all other pixels are relatively de-emphasized to produce the associated output digital values, and (e) binarizing each of said output digital values by associating a first binary value with each of said output digital values which is on one side of a first predetermined threshold value and by associating a second binary value with each of said output digital values which is on the other side of said first predetermined threshold value;

associating a discriminant value with each of said pixels, each discriminant value being chosen so that, for each training object, the dot product of the discriminant values and the binary values is approximately a predetermined nonzero constant value;

repeating steps (a)-(e) for said object; and

comparing the dot product of said discriminant values and the binary values for said object to said predetermined nonzero constant value.

2. The method defined in claim 1 wherein said discriminant values collectively comprise a discriminant image F and wherein said method further comprises the steps of:

converting the binary values associated with said training objects to a plurality of eigenimages Φ

.sup.(j), all of said eigenimages being orthogonal to one another; and

forming said discriminant image from a truncated linear combination of said eigenimages.

3. The method defined in claim 2 wherein the binary values associated with each of said training objects i collectively comprise a training image I.sup.(i), and wherein said method further comprises the step of:

computing each of said eigenimages as in the equation;

##EQU9## where n is the total number of training images and the transformation matrix B is orthogonal.

4. The method defined in claim 3 wherein said step of forming said discriminant image from a truncated linear combination of said eigenimages comprises the step of:

computing said discriminant image as in the equation;

##EQU10## where s is an integer less than n.

5. The method defined in claim 4 further comprising the step of:

selecting the value of s for the equation in claim 4 which gives the minimum value for V(s), where V(s) is computed as in the equation;

##EQU11## where a_i is a nonzero constant for all good training images and zero for any bad training images.

6. The method defined in claim 5 further comprising the step of:

computing the values C_j for the equation of claim 4 as in the equation;

##EQU12## where the eigenvalues γ

_j are the diagonal elements of a matrix G given by the equation AB=BG, where A is the real, symmetric matrix given by the dot product of every possible training image pair, and the transformation matrix B is such that BB⁺ =B⁺ B=1.

7. The method defined in claim 6 wherein said eigenimages are ordered such that γ

₁ >

γ

₂ >

γ

₃ >

. . . >

γ

_n.

8. The method defined in claim 1 wherein said discriminant values are additionally chosen so that the sum of the squares of the discriminant values is approximately minimized.

9. The method defined in claim 1 wherein said discriminant values are additionally chosen so that the sum of the discriminant values is approximately zero.

10. The method defined in claim 1 wherein said two-dimensional images of said training objects may not be in registration with one another, and wherein, prior to said step of associating a discriminant value with each of said pixels, said method further comprises the steps of:

for each training object, (f) computing the center of gravity of at least a portion of the associated two-dimensional image, and (g) shifting the associated binary values so that the centers of gravity associated with all of said training objects are approximately coincident.

11. The method defined in claim 10 wherein, after repeating steps (a)-(e) for said object, said method further comprises the steps of:

computing the center of gravity of at least a portion of the two-dimensional image associated with said object; and

shifting the binary values associated with said object so that the center of gravity associated with said object is approximately coincident with the approximately coincident centers of gravity associated with said training objects.

12. The method defined in claim 1 wherein a plurality of sample objects representative of the range of variation likely to occur among acceptable objects is available, and wherein, prior to said step of repeating steps (a)-(e) for said object, said method further comprises the steps of:

repeating steps (a)-(e) for each sample object;

computing a correlation value for each sample object by forming the dot product of said discriminant values and the binary values for said sample object;

computing the average of the correlation values for all of said sample objects;

computing the standard deviation of the correlation values for all of said sample objects; and

computing first and second threshold correlation values respectively equal to said average of the correlation values plus or minus predetermined numbers of said standard deviation.

13. The method defined in claim 12 wherein said step of comparing the dot product of said discriminant values and the binary values for said object to said predetermined nonzero constant value comprises the step of:

determining whether or not said dot product of said discriminant values and the binary values for said object is between said first and second threshold correlation values.

14. The method defined in claim 1 wherein at least one unacceptable object is available, and wherein said step of associating a discriminant value with each of said pixels comprises the steps of:

repeating steps (a)-(e) for said unacceptable object; and

additionally choosing said discriminant values so that the dot product of the discriminant values and the binary values associated with said unacceptable object is approximately zero.

15. The method defined in claim 1 wherein the number of pixels is substantially greater than the number of training objects.

16. Apparatus for determining whether or not a surface of an object has substantially the same appearance as the corresponding surface of a plurality of training objects comprising:

means for, for each training object, (a) forming a two-dimensional image of said surface, (b) subdividing said two-dimensional image into a plurality of pixels, (c) digitizing each of said pixels by associating with each pixel an initial digital value proportional to an optical characteristic of the associated pixel, (d) edge detecting said initial digital values by processing said initial digital values to produce output digital values such that initial digital values associated with pixels at or near substantial changes in initial digital value are relatively emphasized to produce the associated output digital values, while initial digital values associated with all other pixels are relatively de-emphasized to produce the associated output digital values, and (e) binarizing each of said output digital values by associating a first binary value with each of said output digital values which is on one side of a first predetermined threshold value and by associating a second binary value with each of said output digital values which is on the other side of said first predetermined threshold value;

means for associating a discriminant value with each of said pixels, each discriminant value being chosen so that, for each training object, the dot product of the discriminant values and the binary values is approximately a predetermined non-zero constant value;

means for repeating functions (a)-(e) with respect to said object; and

means for comparing the dot product of said discriminant values and the binary values for said object to said predetermined nonzero constant value.

17. The apparatus defined in claim 16 wherein said discriminant values collectively comprise a discriminant image F and wherein said apparatus further comprises:

means for converting the binary values associated with said training objects to a plurality of eigenimages Φ

.sup.(j), all of said eigenimages being orthogonal to one another; and

means for forming said discriminant image from a truncated linear combination of said eigenimages.

18. The apparatus defined in claim 17 wherein the binary values associated with each of said training objects i collectively comprise a training image I.sup.(i), and wherein said apparatus further comprises:

means for computing each of said eigenimages as in the equation;

##EQU13## where n is the total number of training images and the transformation matrix B is orthogonal.

19. The apparatus defined in claim 18 wherein said means for forming said discriminant image from a truncated linear combination of said eigenimages comprises:

means for computing said discriminant image as in the equation;

##EQU14## where s is an integer less than n.

20. The apparatus defined in claim 19 further comprising:

selecting the value of s for the equation of claim 19 which gives the minimum value for V(s), where V(s) is computed as in the equation;

##EQU15## where a_i is a nonzero constant for all good training images and zero for any bad training images.

21. The apparatus defined in claim 20 further comprising:

means for computing the values C_j for the equation of claim 19 as in the equation;

##EQU16## where the eigenvalues γ

_j are diagonal elements of a matrix G given by the equation AB=BG, where A is the real, symmetric matrix given by the dot product of every possible training image pair, and the transformation matrix B is such that BB⁺ =B⁺ B =1.

22. The apparatus defined in claim 21 wherein said eigenimages are ordered such that γ

₁ >

γ

₂ >

γ

₃ >

. . . >

γ

_n.

23. The apparatus defined in claim 16 wherein said discriminant values are additionally chosen so that the sum of the squares of the discriminant values is approximately minimized.

24. The apparatus defined in claim 16 wherein said discriminant values are additionally chosen so that the sum of the discriminant values is approximately zero.

25. The apparatus defined in claim 16 wherein said two-dimensional images of said training objects may not be in registration with one another, and wherein, prior to operation of said means for associating a discriminant value with each of said pixels, said apparatus further comprises:

means for, for each training object, (f) computing the center of gravity of at least a portion of the associated two-dimensional image, and (g) shifting the associated binary values so that the centers of gravity associated with all of said training objects are approximately coincident.

26. The apparatus defined in claim 25 wherein said means for repeating functions (a)-(e) with respect to said object comprises:

means for computing the center of gravity of at least a portion of the two-dimensional image associated with said object; and

means for shifting the binary values associated with said object so that the center of gravity associated with said object is approximately coincident with the approximately coincident centers of gravity associated with said training objects.

27. The apparatus defined in claim 16 wherein a plurality of sample objects representative of the range of variation likely to occur among acceptable objects is available, and wherein, prior to operation of said means for repeating functions (a)-(e) with respect to said object, said apparatus further comprises the steps of:

means for repeating functions (a)-(e) with respect to each sample object;

means for computing a correlation value for each sample object by forming the dot product of said discriminant values and the binary values for said sample object;

means for computing the average of the correlation values for all of said sample objects;

means for computing the standard deviation of the correlation values for all of said sample objects; and

means for computing first and second threshold correlation values respectively equal to said average of the correlation values plus or minus predetermined numbers of said standard deviation.

28. The apparatus defined in claim 27 wherein said means for comparing the dot product of said discriminant values and the binary values for said object to said predetermined nonzero constant value comprises:

means for determining whether or not said dot product of said discriminant values and the binary values for said object is between said first and second threshold correlation values.

29. The apparatus defined in claim 16 wherein at least one unacceptable object is available, and wherein said means for associating a discriminant value with each of said pixels comprises:

means for repeating functions (a)-(e) with respect to said unacceptable object; and

means for additionally choosing said discriminant values so that the dot product of the discriminant values and the binary values associated with said unacceptable object is approximately zero.

30. The apparatus defined in claim 16 wherein the number of pixels is substantially greater than the number of training objects.

31. The method of determining whether or not an image to be inspected has substantially the same appearance as a plurality of training images, said method comprising the steps of:

digitizing each training image i to produce corresponding training image data I.sup.(i) ;

converting said training image data to a plurality of orthogonal eigenimages given by the equation;

##EQU17## where n is the total number of training images and the transformation matrix B is orthogonal;

computing a discriminant image F from a truncated linear combination of said eigenimages given by the equation;

##EQU18## where s is an integer less than n selected as a value which gives the minimum value for V(s) in the equation;

##EQU19## where a_i is a nonzero constant for all good training images and zero for any bad training images, and where the values C_j are given by the equation;

##EQU20## where the eigenvalues γ

_j are the diagonal elements of a matrix G given by the equation AB=BG, where A is the real, symmetric matrix given by the dot product of every possible training image pair, and the transformation matrix B, is such that BB⁺ =B⁺ B=1;

digitizing said image to be inspected;

forming the dot product of said discriminant image and said digitized image to be inspected; and

comparing said dot product to a predetermined reference value.

32. The method defined in claim 31 wherein said step of computing said discriminant image further comprises the step of:

selecting the individual values which make up said discriminant image so that the sum of the squares of all of said individual values is approximately minimized.

33. The method defined in claim 31 wherein said step of computing said discriminant image further comprises the step of:

selecting the individual values which make up said discriminant image so that the sum of all of said individual values is approximately zero.

34. The method defined in claim 31 wherein said eigenimages are ordered such that γ

₁ >

γ

₂ γ

₃ >

. . . >

γ

_n.