Object recognition using binary image quantization and Hough kernels

• US 7,283,645 B2
• Filed: 06/28/2004
• Issued: 10/16/2007
• Est. Priority Date: 04/13/2000
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Krumm, John, Campbell, Richard
• Primary Examiner(s): Bella; Matthew C.
• Assistant Examiner(s): PERUNGAVOOR, SATHYANARAYA V
• Application Number:
  US10/880,020
• Publication Number:
  US 20040252882A1
• Time in Patent Office: 1,205 Days
• Field of Search: 382/103
• US Class Current: 382/103
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06V 10/757 Matching configurations of points or features

Abstract

A system and process for recognizing an object in an input image involving first generating training images depicting the object. A set of prototype edge features is created that collectively represent the edge pixel patterns encountered within a sub-window centered on each pixel depicting an edge of the object in the training images. Next, a Hough kernel is defined for each prototype edge feature in the form of a set of offset vectors representing the distance and direction, from each edge pixel having an associated sub-window exhibiting an edge pixel pattern best represented by the prototype edge feature, to a prescribed reference point on a surface of the object. The offset vectors are represented as originating at a central point of the kernel. For each edge pixel in the input image, the prototype edge feature which best represents the edge pixel pattern exhibited within the sub-window centered on the edge pixel is identified. Then, for each input image pixel location, the number of offset vectors terminating at that location from Hough kernels centered on each edge pixel location of the input image is identified. The Hough kernel centered on each pixel location is the Hough kernel associated with the prototype edge feature best representing the edge pixel pattern exhibited within a sub-window centered on that input image edge pixel location. The object is declared to be present in the input image if any of the input image pixel locations have a quantity of offset vectors terminating thereat that equals or exceeds a detection threshold.

Claims

1. A system for training at least one general purpose computing device to recognize objects in an input image of a scene, comprising:

at least one general purpose computing device; and

a computer program comprising program modules executable by the at least one computing device, wherein the at least one computing device is directed by the program modules of the computer program to,

generate training images depicting a surface of interest of an object in said input image of the scene,

create a set of prototype edge features which collectively represent the edge pixel patterns encountered within a sub-window centered on each pixel depicting an edge of the object in the training images, and

define a Hough kernel for each prototype edge feature, wherein a Hough kernel for a particular prototype edge feature is defined by a set of offset vectors representing the distance and direction, from each edge pixel having a sub-window associated therewith that has an edge pixel pattern best represented by the prototype edge feature, to a prescribed reference point on the surface of interest of the object, wherein said offset vectors are represented in the Hough kernel as originating at a central point thereof, wherein

said set of prototype edge features and said Hough kernels associated therewith are used to recognize said object in the input image and to identify the location of the object in the input image.

2. The system of claim 1, wherein the program module for generating training images, comprises sub-modules for:

inputting a model image that depicts the surface of interest of the object;

creating a base training image depicting the portion of the model image containing said surface of interest; and

synthesizing training images from the base image which depict the object'"'"'s surface of interest in different orientations.

3. The system of claim 2, wherein the sub-module for inputting a model image, comprises a sub-module for inputting a model image that was captured from a viewpoint where the normal vector of the object'"'"'s surface of interest is pointed at the camera so as to be substantially coincident with the optical axis of the camera.

4. The system of claim 2, wherein the sub-module for creating a base training image, comprises sub-modules for:

identifying the surface of interest in the model image; and

extracting the portion of the model image depicting the object'"'"'s surface of interest.

5. The system of claim 4, wherein the sub-module for identifying the surface of interest in the model image, comprises a sub-module for allowing a user to manually select the surface of interest in the model image.

6. The system of claim 5, wherein the sub-module for allowing the user to select the surface of interest in the model image, comprises sub-modules for:

displaying the model image to the user on a computer monitor screen; and

requesting the user to outline the object'"'"'s surface of interest on the display of the model image.

7. The system of claim 2, wherein the model image, base training image, and synthetic training images are color images.