Method and apparatus for segmenting images using constant-time deformable contours

• US 6,031,935 A
• Filed: 02/12/1998
• Issued: 02/29/2000
• Est. Priority Date: 02/12/1998
• Status: Expired due to Fees

• Current Assignee: Zebadiah M. Kimmel
• Original Assignee: Zebadiah M. Kimmel
• Inventors: Kimmel, Zebadiah M.
• Primary Examiner(s): Tran, Phuoc
• Application Number:
  US09/022,870
• Time in Patent Office: 747 Days
• Field of Search: 382/173, 382/170, 382/199, 382/203, 382/286, 382/288, 382/295, 382/296
• US Class Current: 382/173
• CPC Class Codes: G06T 2207/10088 Magnetic resonance imaging [MRI]
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 2207/30016 Brain
• CPC Class Codes: G06T 7/0012 Biomedical image inspection
• CPC Class Codes: G06T 7/12 Edge-based segmentation
• CPC Class Codes: G06T 7/149 involving deformable models, e.g. active contour models
• CPC Class Codes: G06V 10/46 Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features colour feature extraction G06V10/56

Abstract

An image segmentation system uses deformable contours to automatically identify the border of a target object in each image of a set of images. The majority of contour deformations, or moves, are performed in O(1) time. Each contour uses both global and local geometric information and pixel intensity information extracted from a training contour, as well as a priori domain knowledge, to find the outline of the target object.

Claims

1. A method of segmenting a set of images comprising the steps of:

a) providing a set of search images, each search image containing a view of a target object;

b) providing a priori domain knowledge corresponding to the set of search images;

c) selecting a training image from the set of search images;

d) obtaining at least one training contour from the selected training image, the training contour representing an outline of the target object in the selected training image;

e) utilizing the selected training contour to extract training information corresponding to the target object in the selected training image;

f) associating the training contour with a selected search image of the set of images, the training contour defining a search contour;

g) deforming the search contour relative to the selected search image and maintaining predetermined information corresponding to the deformed search contour;

h) utilizing the training information and the a priori domain knowledge to identify regions of the deformed search contour that correspond within predetermined limits to an outline of the target object in the selected search image; and

i) repeating steps (c) through (h) for each image in the set of search images so as to segment all of the images in the set of search images.

2. The method according to claim 1 wherein the step of providing a priori domain information includes providing information in the form of predetermined rules relating to known characteristics corresponding to the search images.

3. The method according to claim 1 wherein at least one of the training contour and the search contour is represented as a plurality of cracks in a list of cracks, each crack representing the space between adjacent pixels of the target object and having an assigned value equal to one of four discrete directions, the direction assigned to each crack indicating a location corresponding to an adjacent crack of the contour, such that the list of cracks define a closed, non-intersecting contour having an interior portion and an exterior portion.

4. The method according to claim 3 wherein the assigned directions "d" of a crack are rotated utilizing the following formulas, said assigned directions being expressed as hexadecimal values of 0xF0, 0xD0, 0xC0, and 0xE0, respectively, the formulas for rotation defined as:

1) Rightward rotation of a direction of crack by ninety degrees=

space="preserve" listing-type="equation">(d-0x10)|0xC0;

2) Leftward rotation of a direction of a crack ninety degrees=

space="preserve" listing-type="equation">(d+0x10)|0xC0;

3) One-hundred and eighty degree rotation of a direction of a crack=

space="preserve" listing-type="equation">(d 0x20); and

where the symbol "|" is equal to the bitwise logical OR operation and the symbol " " is equal to the bitwise logical exclusive OR (XOR) operation.

5. The method according to claim 1 wherein at least one of the training contour and the search contour is partitioned into a plurality of segments defined by a portion of the contour subtended by an arc formed by rays originating at a centroid of the contour, each arc having equal angular values.

6. The method according to claim 5 wherein each segment has a corresponding histogram representing pixel characteristics of the segment, the pixel characteristics represented as pixel intensity pairs stored in the histogram, each pixel intensity pair corresponding to the intensity of at least one pixel located in the interior portion of the contour and the intensity of at least one pixel located in the exterior portion of the contour, respectively.

7. The method according to claim 6 wherein each pixel intensity pair is generated by averaging values of pixels located within a predetermined distance from the contour and lying along a radius of the contour as defined by its centroid.

8. The method according to claim 6 wherein each pixel intensity pair is used to place related pixel intensity pairs in the histogram, so that the original pixel intensity pair is blurred.

9. The method according to claim 8 wherein each pixel intensity pair is used to place other pixel intensity pairs in the histogram, such that a totality of pairs define a circle centered at the original pixel intensity pair.

10. The method according to claim 6 wherein the histograms corresponding to the search contour are formed as chained hash tables so as to minimize an amount of memory required to contain the pixel intensity pair data for the contour.

11. The method according to claim 6 wherein the pixel intensity pairs corresponding to a plurality of pixels of the search image, relative to the centroid of the training contour, are stored in a look-up table prior to the step of identifying regions of the deformed search contour.

12. The method according to claim 1 wherein the step of deforming the search contour includes the step of performing at least one of a plurality of simple moves and a plurality of complex moves, said moves modifying at least one crack in a list of cracks that represent the search contour.

13. The method according to claim 12 wherein the simple move displaces at least one crack of the search contour from its original position by one crack position, the simple move affecting a maximum of four cracks in the crack list, said simple move occurring in O(1) order of time if the simple move does not shift the centroid of the search contour by more than a predetermined distance.

14. The method according to claim 12 wherein predetermined information corresponding to the search contour is updated after the search contour is deformed, said updating requiring order O(1) time to perform.

15. The method according to claim 12 wherein if the deformation of the search contour causes a centroid of the search contour to be displaced by less than a predetermined number of pixels, the predetermined information corresponding to the search contour is updated and requires order O(1) time to perform the update.

16. The method according to claim 12 wherein if the deformation of the search contour causes a centroid of the search contour to be displaced by more than a predetermined number of pixels, the predetermined information corresponding to the search contour is updated and requires up to order O(n) time to perform the update, and where n is equal to the number of cracks in the search contour.

17. The method according to claim 6 wherein the step of identifying regions of the deformed search contour that correspond to the outline of the target object include the step of determining the amount of overlap between the training contour and the search contour respectively.

18. The method according to claim 17 wherein the step of deforming the search contour is performed by growing and shrinking the search contour according to a predetermined sequence such that deformations that increase the amount of overlap between the training contour and the search contour are accepted, and deformations that decrease the amount of overlap between the training contour and the search contour are rejected.

19. The method according to claim 17 wherein the step of alternating between growing and shrinking permits the deformed search contour to jump out of local minimum conditions.

20. The method according to claim 3 wherein after the search contour is deformed, if a non-closed contour results, cracks within the list of cracks are connected to form a closed non-intersecting contour.

21. The method according to claim 1 wherein multiple training images are utilized to extract the training information.

22. The method according to claim 1 wherein the training contour is provided by a human operator.

23. The method according to claim 1 wherein the training contour is supplied by a predetermined library of training contours corresponding to the set of search images.

24. A method of segmenting a set of images comprising the steps of

a) providing a set of search images, each search image containing a view of a target object;

b) providing a priori domain knowledge corresponding to the set of search images;

c) selecting a training image from the set of search images;

d) obtaining a training contour from the selected training image, the training contour representing an outline of the target object in the selected training image;

e) utilizing the selected training contour to extract training information corresponding to the target object in the selected training image;

f) associating the training contour with a selected search image of the set of images, the training contour defining a search contour;

g) deforming the search contour relative to the selected search image and maintaining predetermined information corresponding to the deformed search contour;

h) utilizing the training information and the a priori domain knowledge to identify regions of the deformed search contour that correspond within predetermined limits to an outline of the target object in the selected search image;

i) utilizing selected regions of the search contour and the a priori domain knowledge to fashion a final segmentation contour such that the final segmentation contour forms an outline of the target object in the selected search image; and

j) repeating steps (c) through (i) for each image in the set of search images so as to segment all of the images in the set of search images.