Method and system for fuzzy clustering of images

• US 6,798,911 B1
• Filed: 03/28/2001
• Issued: 09/28/2004
• Est. Priority Date: 03/28/2001
• Status: Expired due to Term

• Current Assignee: AT&T Corporation (AT&T, Inc.)
• Original Assignee: AT&T Corporation (AT&T, Inc.)
• Inventors: Jafarkhani, Hamid, Tarokh, Vahid
• Primary Examiner(s): Bali, Vikkram
• Application Number:
  US09/819,557
• Time in Patent Office: 1,280 Days
• Field of Search: 382/115-128, 382/181, 382/224, 382/225, 358/3.17, 704/245, 348/700
• US Class Current: 382/225
• CPC Class Codes: G06F 18/2323 based on graph theory, e.g. minimum spanning trees [MST] or graph cuts

Abstract

An approach to clustering a set of images based on similarity measures employs a fuzzy clustering paradigm in which each image is represented by a node in a graph. The graph is ultimately partitioned into subgraphs, each of which represent true clusters among which the various images are distributed. The partitioning is performed in a series of stages by identifying one true cluster at each stage, and removing the nodes belonging to each identified true cluster from further consideration so that the remaining, unclustered nodes may then be grouped. At the beginning of each such stage, the nodes that remain to be clustered are treated as all belonging to a single candidate cluster. Nodes are removed from this single candidate cluster in accordance with similarity and connectivity criteria, to arrive at a true cluster. The member nodes of this true cluster are then removed from further consideration, prior to the next stage in the process.

Claims

1. A method for clustering a set of N images into P final clusters, where N and P are integers, the method comprising the steps of:

(a) calculating at least one similarity measure S_ij between members of each possible pair of images, wherein S_ij represents the similarity measure between the ith and jth images with i and j being image indices and S_ij=S_j,i;

(b) calculating a total connectivity value for each of the images remaining to be clustered, a total connectivity value for each image being defined as a sum of a function f of the similarity measures associated with that image;

(c) identifying, from among said images remaining to be clustered, a maximum total connectivity value T_max corresponding to an image I_max, image I_max belonging to a current cluster C which initially includes all images remaining to be clustered;

(d) removing, from the current cluster C, at least one image based on at least one of its similarity measure with image I_max and its total connectivity value within current cluster C;

(e) adding, to current cluster C, images having a similarity measure that is greater than a threshold T3 with any image currently in C;

(f) calculating, for each image within current cluster C, a total connectivity value based on those images within C;

(g) removing, from current cluster C, those images having a total connectivity value less than some threshold T4;

(h) repeating steps (f) and (g) until no further images are removed to thereby establish current cluster C as one of the final clusters;

(i) removing all images in current cluster C from further consideration; and

(j) repeating steps (b)-(i) until all N images are assigned to a final cluster.

2. The method according to claim 1, comprising the step of applying a non-linear function to each similarity measure, prior to calculating the total connectivity value.

3. The method according to claim 2, wherein the non-linear function is a transcendental function.

4. The method according to claim 3, wherein the transcendental function is a hyperbolic tangent.

5. The method according to claim 4, wherein each similarity measure S_ij is transformed according to the formula tanh(5×

S_ij), prior to calculating a total connectivity value.

6. The method according to claim 1, wherein the function f is the identify function.

7. The method according to claim 1, wherein the function f is a nonlinear function.

8. The method according to claim 1, wherein in step (e), at least one image is removed from the current cluster based on its similarity measure with image I_max, and then at least one other image is removed from the current cluster based on its total connectivity value within the current cluster.

9. The method according to claim 1, wherein in step (e), at least one image is removed from the current cluster based on its total connectivity value within the current cluster, and then at least one other image is removed from the current cluster based on its similarity measure with image I_max.

10. The method according to claim 1, wherein in step (e), images having a similarity measure with I_max that is less than some threshold T1 are removed from the current cluster.

11. The method according to claim 1, wherein in step (e), a number N1 images having the N1 lowest similarity measures with I_max are removed from the current cluster.

12. The method according to claim 1, wherein in step (e), images having a total connectivity value which differs from t_max by more than some threshold T2 are removed from the current cluster.

13. The method according to claim 12, wherein images having a similarity measure with I_max that is less than some threshold T1 are removed from the current cluster.

14. The method according to claim 13, wherein at least one image is removed from the current cluster based on its similarity measure with image I_max, and then at least one other image is removed from the current cluster based on its total connectivity value within the current cluster.

15. The method according to claim 12, wherein a number N1 images having the N1 lowest similarity measures with I_max are removed from the current cluster.

16. The method according to claim 15, wherein at least one image is removed from the current cluster based on its similarity measure with image I_max, and then at least one other image is removed from the current cluster based on its total connectivity value within current cluster.

17. A method for clustering a set of N items into P final clusters, where N and P are integers, the method comprising the steps of:

(a) calculating at least one similarity measure S_ij between members of each possible pair of items, wherein S_ij represents the similarity measure between the ith and jth items with i and j being item indices and S_ij=S_j,i;

(b) calculating a total connectivity value for each of the items remaining to be clustered, a total connectivity value for each item being defined as a sum of a function f of the similarity measures associated with that item;

(c) identifying, from among said items remaining to be clustered, a maximum total connectivity value t_max corresponding to an item I_max, item I_max belonging to a current cluster C which initially includes all items remaining to be clustered;

(d) removing, from the current cluster C, at least one item based on at least one of its similarity measure with item I_max and its total connectivity value within current cluster C;

(e) adding, to current cluster C, items having a similarity measure that is greater than a threshold T3 with any item currently in C;

(f) calculating, for each item within current cluster C, a total connectivity value based on those items within C;

(g) removing, from current cluster C, those items having a total connectivity value less than some threshold T4;

(h) repeating steps (f) and (g) until no further items are removed to thereby establish current cluster C as one of the final clusters;

(i) removing all items in current cluster C from further consideration; and

(j) repeating steps (b)—

(i) until all N items are assigned to a final cluster.