Machine learning apparatus and method for image searching

• US 5,793,888 A
• Filed: 01/29/1997
• Issued: 08/11/1998
• Est. Priority Date: 11/14/1994
• Status: Expired due to Fees

• Current Assignee: Massachusetts Institute of Technology
• Original Assignee: Massachusetts Institute of Technology
• Inventors: Delanoy, Richard L.
• Primary Examiner(s): Mancuso, Joseph
• Assistant Examiner(s): Johnson, Timothy M.
• Application Number:
  US08/790,178
• Time in Patent Office: 559 Days
• Field of Search: 382/305, 382/209, 382/155, 382/215, 382/216, 382/219, 382/222, 382/170, 382/168, 382/141, 382/217, 364/419.19
• US Class Current: 382/219
• CPC Class Codes: G06F 16/5838 using colour
• CPC Class Codes: G06F 16/5854 using shape and object relationship
• CPC Class Codes: G06F 18/40 Software arrangements specially adapted for pattern recognition, e.g. user interfaces or toolboxes therefor
• CPC Class Codes: G06V 20/13 Satellite images

Abstract

A computer module is interactively trained to recognize patterns of spectral or textural features in imagery. The interactively trained module is based on techniques of knowledge-based image processing and an approach that uses interest images to provide a means of continuous feedback of module performance to the user. The user, in turn, responds by indicating where the module is making mistakes. A set of user indicated examples and counter examples form the inputs to a machine learning program called Functional Template Learning. The trained module is exportable as an independent agent to search large databases for matching patterns. Following retrieval of images having matching patterns, a user can further refine the agent'"'"'s performance by indicating where mistakes have been made. The overall search tool employing the trained module is capable of prescreening and highlighting images, significantly reducing the workload of analysts attempting to detect regions or objects in imagery.

Claims

1. Computer image searching apparatus for obtaining a desired image from a multiplicity of images, comprising:

input means for setting forth a user-desired image subject to be recognized, the image subject being formed of a plurality of attributes;

processor means coupled to the input means for determining from the plurality of attributes a subset of most discriminating attributes, the processor means including;

a histogram member coupled to the input means for generating a pair of histograms for each attribute, the pair including a first histogram based on user-input examples of the attribute and a second histogram based on user-input counter examples of the attribute;

processing means coupled to the histogram member for generating for each attribute, a weight and a scoring function from the pair of histograms generated for the attribute; and

a sorter for selecting scoring functions according to highest weight, the attributes corresponding to said selected scoring functions forming the subset of most discriminating attributes;

a functional template member coupled to the processor means for providing a functional template based on scoring functions for each discriminating attribute in the subset; and

an agent coupled to the functional template member to receive the functional template therefrom, the agent employing the functional template in a matching routine to extract images from the multiplicity of images, the extracted images having the desired image subject.

2. Computer image searching apparatus as claimed in claim 1 wherein for each attribute, the processing means calculates the weight as a function of degree of overlap between the first and second histograms in the pair of histograms for the attribute.

3. Computer image searching method for obtaining a desired image from a multiplicity of images, comprising the steps of:

(a) providing a user-desired image subject to be recognized, the image subject being formed of a plurality of attributes;

(b) using processor means, determining from the plurality of attributes, a subset of most discriminating attributes by;

generating for each attribute a pair of histograms, the pair including a first histogram based on user input examples of the attribute and a second histogram based on user input counter examples of the attributes;

generating weights and scoring functions of the attributes based on the histograms, including utilizing the pair of histograms generated for the attribute; and

selecting scoring functions according to highest weight, the attributes corresponding to said selected scoring functions forming the subset of most discriminating attributes;

(c) generating a functional template based on scoring functions for each discriminating attribute in the subset; and

(d) using the functional template, matching a model of the desired image subject to images in the multiplicity of images, to extract those images having the desired image subject.

4. Computer image searching method as claimed in claim 3 wherein the step of generating weights and scoring functions of the attributes includes calculating the weight as a function of degree of overlap between the first and second histograms in the pair of histograms for each attribute.

5. Computer image searching apparatus for obtaining a desired image from a multiplicity of images, comprising:

input means for setting forth a user-desired image subject to be recognized, the image subject being formed of a plurality of attributes;

processor means coupled to the input means for determining from the plurality of attributes a subset of most discriminating attributes, the processor means including;

a histogram member coupled to the input means for generating a histogram for each attribute;

processing means coupled to the histogram member for generating weights and scoring functions for the attributes; and

a sorter for selecting scoring functions according to highest weight, the attributes corresponding to said selected scoring functions forming the subset of most discriminating attributes;

wherein the processing means and sorter are coupled such that the processing means generates weights for each attribute, then the sorter selects attributes with the highest weights, and the processing means generates scoring functions for the selected attributes;

a functional template member coupled to the processor means for providing a functional template based on scoring functions for each discriminating attribute in the subset; and

an agent coupled to the functional template member to receive the functional template therefrom, the agent employing the functional template in a matching routine to extract images from the multiplicity of images, the extracted images having the desired image subject.

6. Computer image searching method for obtaining a desired image from a multiplicity of images, comprising the steps of:

(a) providing a user-desired image subject to be recognized, the image subject being formed of a plurality of attributes;

(b) using processor means, determining from the plurality of attributes, a subset of most discriminating attributes by;

generating for each attribute a histogram;

based on the histograms, generating weights for each attribute;

selecting attributes with the highest weights;

generating scoring functions for only the selected attributes; and

selecting scoring functions according to highest weight, the attributes corresponding to said selected scoring functions forming the subset of most discriminating attributes;

(c) generating a functional template based on scoring functions for each discriminating attribute in the subset; and

(d) using the functional template, matching a model of the desired image subject to images in the multiplicity of images, to extract those images having the desired image subject.