Video recognition system

• US 4,611,347 A
• Filed: 09/24/1984
• Issued: 09/09/1986
• Est. Priority Date: 09/24/1984
• Status: Expired due to Term

• Current Assignee: Bell Telephone Laboratories, Inc. (Nokia Corporation)
• Original Assignee: AT&T, Inc.
• Inventors: Netravali, Arun N., Segen, Jakub
• Primary Examiner(s): Boudreau, Leo H.
• Application Number:
  US06/653,213
• Time in Patent Office: 715 Days
• Field of Search: 382/38, 382/37, 382/34, 382/30, 382/27
• US Class Current: 382/159
• CPC Class Codes: G06V 30/2504 Coarse or fine approaches, e.g. resolution of ambiguities or multiscale approaches

Abstract

An image recognition system uses feature or attribute matching for fast coarse selection of a group of reference images which closely resemble an unknown image and thereafter selects the best reference image match from that group using 2-dimensional intensity correlation or template matching. The system provides highly accurate image recognition without resorting to time consuming exhaustive search by template matching.

Claims

1. A method of operating an image recognition system comprising the steps of:

segmenting a video image into two or more regions based on variations in intensity in said video image;

screening the characteristics of each of said regions by comparing curvature and orientation attributes of selected local features of regions with reference features stored in an attribute memory and identifying one or more sets of reference features from said attribute memory which closely resemble a set of selected local features of each region;

retrieving one or more reference images for each regions from an image memory using addresses from the screening and identifying step;

for each of the reference images, computing a horizontal and a vertical offset by matching each of said selected local features with each of said reference features; and

using the offsets computed in the previous step, aligning and matching the intensity of each of said regions against the intensity of each of the retrieved reference images for that region to determine one reference image which most closely matches that region.

2. The method of claim 1 wherein said segmenting step includes an edge following algorithm to identify the regions.

3. The method of claim 1 wherein said segmenting step includes a difference thresholding algorithm to identify the regions.

4. The method of claim 1 wherein said segmenting step includes an image thresholding algorithm to identify the regions.

5. The method of claim 1 wherein said screening and identifying step uses a hierarchical classification method for identifying sets of reference features.

6. The method of claim 1 wherein said aligning and matching step uses a two dimensional normalized correlation to determine the degree of match between the reference image and the region.

7. The method of claim 1 wherein said aligning and matching step uses an occluded image matching algorithm to determine the degree of match between a reference image and an occluded object in that region.

8. The method of claim 1 wherein the image memory has fast sequential access.

9. The method of claim 8 wherein the image memory has slow random access.

10. The method of claim 9 wherien the image memory is a video disc.

11. An image recognition system comprising

means for segmenting a video image into two or more regions based on variations in intensity in said video image;

means for screening the characteristics of each of said regions by comparing curvature and orientation attributes of selected local features of regions with reference features stored in an attribute memory and identifying one or more sets of reference features from said attribute memory which closely resemble a set of selected local features of each region;

means for retrieving one or more reference images for each region from an image memory using addresses from the screening and identifying step;

means for computing for each of the reference images a horizontal and a vertical offset by matching each of said selected local features with each of said reference features; and

means for aligning using the offsets computed by said computing means and for matching the intensity of each of said regions against the intensity of each of the retrieved reference images for that region to determine one reference image which most closely matches that region.