SYSTEM AND METHOD FOR HIERARCHICAL IMAGE PROCESSING

• US 20110055266A1
• Filed: 08/25/2009
• Published: 03/03/2011
• Est. Priority Date: 08/25/2009
• Status: Active Grant

• Current Assignee: Tech Mahindra Limited
• Original Assignee: Satyam Computer Services Limited (Tech Mahindra Limited)
• Inventors: Aleti, Adhipathi Reddy, Gangadharpalli, Sridhar, Varadarajan, Sridhar
• Granted Patent: US 8,068,677 B2
• Time in Patent Office: Days
• US Class Current: 707/780
• CPC Class Codes: G06F 16/583 using metadata automatically derived from the content

Abstract

Video processing in general and image processing in particular is a very complex process. This is more so when the analysis is required to extract information from the images for providing of the same directly or indirectly to the users. This additional complexity is due to the direct competition with users who are used to analyzing of the images both syntactically and semantically. One of the ways to contain the complexity of image analysis is to exploit the domain semantics during image processing. A system and method to help in semantics based image processing involves the identification of one or more domain relevant semantic hierarchies and using of the same during image processing.

Claims

1. A method for improving the recognition accuracy of a plurality of objects possibly contained in an image by performing hierarchical image processing to determine a plurality of image annotations of said image, wherein said method comprising:

obtaining of a plurality of classifiers, wherein each of said plurality of classifiers is associated with an object of said plurality of objects;

obtaining of a plurality of feature vectors, wherein each of said plurality of feature vectors is associated with a class of said plurality of classifiers;

obtaining of a plurality of labels, wherein each of said plurality of labels is associated with a classifier of said plurality of classifiers;

obtaining of a plurality of plurality of semantic relationships, wherein each of said plurality of plurality of semantic relationships is associated with a label of said plurality of labels;

computing of a plurality of factors associated with a plurality of hierarchies, wherein each of said plurality of hierarchies is based on said plurality of plurality of semantic relationships;

computing of a hierarchy measure for said plurality of hierarchies;

determining of a plurality of formed hierarchies based on said plurality of classifiers, said plurality of labels, and said plurality of plurality of semantic relationships;

analyzing of said image based on said plurality of formed hierarchies resulting in a plurality of plurality of hierarchical annotations; and

performing of consistency analysis based on said plurality of plurality of hierarchical annotations resulting in said plurality of image annotations.

2. The method of claim 1, wherein said method of computing of said plurality of factors further comprising:

determining of a plurality of sibling sets based on said plurality of labels, a sibling relationship among said plurality of labels, and said plurality of plurality of semantic relationships, wherein each of said plurality of sibling sets is a plurality of siblings.

determining of a total number of elements in said plurality of sibling sets;

determining of a total number of hierarchies in said plurality of hierarchies;

determining of a plurality of number of hierarchies, wherein each of said plurality of number of hierarchies is based on the number of hierarchies across which a plurality of siblings of a sibling set of said plurality of sibling sets is distributed;

determining of a plurality of number of elements, wherein each of said plurality of number of elements is based on a sibling set of said plurality of sibling sets;

computing of a sibling factor of said plurality of factors based on said total number of elements, said total number of hierarchies;

said plurality of number of hierarchies, and said plurality of number of elements.

3. The method of claim 2, wherein said method further comprising:

determining of a total number of labels in said plurality of labels;

determining of a plurality of number of hierarchies, wherein each of said plurality of number of hierarchies is based on the number of hierarchies across which a label of said plurality of labels is distributed; and

computing of a redundancy factor of said plurality factors based on said total number of labels and said plurality of number of hierarchies.

4. The method of claim 2, wherein said method further comprising:

obtaining of a total number of pairs based on said plurality of labels;

determining of a plurality of near-far values, wherein each of said plurality of near-far values is based on a near-far relationship among said plurality of labels, said plurality of plurality of semantic relationships, and a pair of labels, wherein each of said pair of labels is part of said plurality of labels;

determining of a plurality of near-far hierarchical values, wherein each of said plurality of near-far hierarchical values is based on a pair of labels, wherein each of said pair of labels is part of said plurality of labels, the distance between said pair of labels based on said plurality of hierarchies, and a plurality of maximum path lengths, wherein each of said plurality of maximum path lengths is associated with a hierarchy of said plurality of hierarchies; and

computing of a near-far factor of said plurality of factors based on said plurality of near-far values, said plurality of near-far hierarchical values, and said total number of pairs.

5. The method of claim 2, wherein said method further comprising:

determining of a total number of labels in said plurality of labels;

determining of a plurality of depths, wherein each of said plurality of depths is associated with a depth of a label based on said plurality of hierarchies;

computing of a conflict factor of said plurality of factors based on said total number of labels and said plurality of depths.

6. The method of claim 1, wherein said method of computing of said hierarchy measure further comprising:

computing of a sibling factor based on said plurality of hierarchies;

computing of a redundancy factor based on said plurality of hierarchies;

computing of a near-far factor based on said plurality of hierarchies;

computing of a conflict factor based on said plurality of hierarchies;

obtaining of a weight 1 associated with said sibling factor of said plurality of factors;

obtaining of a weight 2 associated with said redundancy factor of said plurality of factors;

obtaining of a weight 3 associated with said near-far factor of said plurality of factors;

obtaining of a weight 4 associated with said conflict factor of said plurality of factors; and

computing of said hierarchy measure based on said sibling factor, said weight 1, said redundancy factor, said weight 2, said near-far factor, said weight 3, said conflict factor, and said weight 4.

7. The method of claim 1, wherein said method of determining further comprising:

randomly forming of a plurality of hierarchy sets, wherein each of said plurality of hierarchy sets is a plurality of hierarchies based on said plurality of labels and said plurality of plurality of semantic relationships;

computing of a plurality of hierarchy measures, wherein each of said plurality of hierarchy measures is associated with a hierarchy set of said plurality of hierarchy sets; and

applying of a stochastic optimization technique based on said plurality of hierarchy sets and said plurality of hierarchy measures to determine said plurality of formed hierarchies.

8. The method of claim 1, wherein said method of analyzing further comprising:

obtaining of a formed hierarchy of said plurality of formed hierarchies;

obtaining of a node of said formed hierarchy wherein said node is marked for traversal;

obtaining of a plurality of classifiers associated with said node;

applying of each of said plurality of classifiers with respect to said image resulting in a plurality of recognition accuracies;

obtaining of a recognition accuracy of said plurality of accuracies, wherein said recognition accuracy is associated with a classifier of said plurality of classifiers and said recognition accuracy exceeds a pre-defined threshold;

obtaining of a child node of said formed hierarchy associated with said classifier;

marking of said child node for traversal;

obtaining of a label associated with child node; and

making of said label part of a plurality of hierarchical annotations of said plurality of plurality of hierarchical annotations, wherein said plurality of hierarchical annotations is associated with said formed hierarchy.

9. The method of claim 1, wherein said method of performing further comprising:

obtaining of a plurality of hierarchical annotations of said plurality of plurality of hierarchical annotations;

forming of a plurality of plurality of consistent annotations based on said plurality of hierarchical annotations, said plurality of plurality of hierarchical annotations, and said plurality of plurality of semantic relationships;

selecting a plurality of maximal consistent annotations based on said plurality of plurality of consistent annotations, wherein said plurality of maximal consistent annotations is maximal among said plurality of plurality of consistent annotations; and

making of said plurality of maximal consistent annotations as said plurality of image annotations.