Object/anti-object neural network segmentation

• US 5,245,672 A
• Filed: 03/09/1992
• Issued: 09/14/1993
• Est. Priority Date: 03/09/1992
• Status: Expired due to Fees

• Current Assignee: The United States of America As Represented By The Secretary of Agriculture
• Original Assignee: THE GOVERNMENT OF THE UNITED STATES OF AMERICA, AS REPRESENTED BY THE SECRETARY OF COMMERCE
• Inventors: Wilkinson, Robert A. Jr., Garris, Michael D., Wilson, Charles L.
• Primary Examiner(s): Boudreau, Leo H.
• Application Number:
  US07/847,490
• Time in Patent Office: 554 Days
• Field of Search: 382/9, 382/48, 382/14, 382/15, 382/34, 382/32, 382/33
• US Class Current: 382/179
• CPC Class Codes: G06F 18/24317 Piecewise classification, i.e. whereby each classification requires several discriminant rules
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/18057 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 30/414 Extracting the geometrical structure, e.g. layout tree; Block segmentation, e.g. bounding boxes for graphics or text

Abstract

The system of the present invention applies self-organizing and/or supervd learning network methods to the problem of segmentation. The segmenter receives a visual field, implemented as a sliding window and distinguishes occurrences of complete characters from occurrences of parts of neighboring characters. Images of isolated whole characters are true objects and the opposite of true objects are anti-objects, centered on the space between two characters. The window is moved across a line of text producing a sequence of images and the segmentation system distinguishes true objects from anti-objects. Frames classified as anti-objects demarcate character boundaries, and frames classified as true objects represent detected character images. The system of the present invention may be a feedforward adaption using a symmetric triggering network. Inputs to the network are applied directly to the separate associative memories of the network. The associative memories produce a best match pattern output for each part of the input data. The associative memories provide two or more subnetworks which define data subsets, such as objects or anti-objects, according to previously learned examples. Multi-layer perceptron architecture may also be used in the system of the present invention rather than the symmetrically triggered feedforward adaptation with tradeoffs in training time but advantages in speed.

Claims

1. A segmentation network for segmenting an image having features including a plurality of characters and a plurality of negative spaces between said characters, comprising:

input means for receiving an input signal representative of an image to be segmented;

means for extracting said features in response to said input signal;

classifying means for classifying said extracted features;

negative space determining means for determining a negative space of said plurality of negative spaces between said characters; and

,

said classifying means having means for preventing classification in accordance with said negative space determining means.

2. The segmentation network of claim 1, including self-organizing associative memory means comprising means for extracting said features and classifying said extracted features simultaneously.

3. The segmentation network of claim 2, wherein said negative space determining means comprises associative memory anti-object detection means.

4. The segmentation network of claim 3, wherein said associative memory anti-object detection means comprises means for detecting the anti-object of said image.

5. The segmentation network of claim 4, further comprising associative memory object detection means for determining said characters.

6. The segmentation network of claim 5, wherein said object detection means and said anti-object detection means are triggered symmetrically.

7. The segmentation system of claim 5, wherein differing associative memory detection means uses separate matching functions and stores objects and anti-objects in optimum metaclasses.

8. The segmentation system of claim 5, wherein differing associative memory detection means use separate learning rules.

9. The segmentation network of claim 5, wherein said associative memory object detection means and said associative memory object detection means operate in parallel on parallel data.

10. The segmentation network of claim 1, including multi-layer perceptron means wherein said means for extracting said features comprises a filter.

11. The segmentation network of claim 10, wherein said filter comprises a Karhunen-Loeve filter.

12. The segmentation network of claim 10, wherein said filter comprises a Gabor filter.

13. The segmentation network of claim 10, wherein said means for preventing classification comprises negative space gate means for controlling classification in accordance with anti-object detection.

14. The segmentation network of claim 1, wherein said extracted features correspond to both said characters and said negative spaces.

15. The segmentation network of claim 1, wherein all negative spaces of said plurality of negative spaces are represented as a single generalized negative space.