Neural network based character position detector for use in optical character recognition

• US 5,542,006 A
• Filed: 06/21/1994
• Issued: 07/30/1996
• Est. Priority Date: 06/21/1994
• Status: Expired due to Term

• Current Assignee: Ip Ot Sub Ulc
• Original Assignee: Eastman Kodak Company
• Inventors: Shustorovich, Alexander, Thrasher, Christopher W.
• Primary Examiner(s): Mancuso, Joseph
• Application Number:
  US08/263,304
• Time in Patent Office: 770 Days
• Field of Search: 395/20, 395/21, 395/22, 395/23, 395/24, 382/155, 382/156, 382/161, 382/157, 382/158, 382/159, 382/186, 382/187, 382/173
• US Class Current: 382/156
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/146 Aligning or centring of the image pick-up or image-field
• CPC Class Codes: G06V 30/153 using recognition of characters or words
• CPC Class Codes: G06V 30/18057 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

Apparatus, and an accompanying method, for use in an optical character recognition (OCR) system (5) for locating, e.g., center positions ("hearts") of all desired characters within a field (310; 510) of characters such that the desired characters can be subsequently recognized using an appropriate classification process. Specifically, a window (520) is slid in a step-wise convolutional-like fashion (520₁, 520₂, 520₃) across a field of preprocessed, specifically uniformly scaled, characters. Each pixel in the window is applied as an input to a positioning neural network (152) that has been trained to produce an output activation whenever a character "heart" is spatially coincident with a pixel position within an array (430) centrally located within the window. As the window is successively moved across the field, in a stepped fashion, the activation outputs of the neural network are averaged, on a weighted basis, for each different window position and separately for each horizontal pixel position in the field. The resulting averaged activation output values, typically in the form of a Gaussian distribution for each character, are then filtered, thresholded and then used, via a weighted average calculation with horizontal pixel positions being used as the weights, to determine the character "heart" position as being the center pixel position in the distribution.

Claims

1. An optical character recognition system for recognizing a character situated within an image field, wherein the image field contains image pixels and said system detects a position of a pre-defined element of said character within the image field, said system comprising:

means for successively extracting pre-defined array portions from said field so as to define a plurality of corresponding windows, all of said windows having a pre-defined size, wherein each of the windows has a starting position relative to the field that is offset, along a horizontal field direction, by a pre-defined pixel amount with respect to the starting position of an adjacent one of the windows relative to the field;

a first neural network, responsive to said extracting means and having a plurality of first input nodes and a plurality of first output nodes, wherein each output node is associated with a different horizontal pixel location within an array portion of each window, and the first neural network;

generates an activation output at one of the first output nodes whenever the position of the pre-defined element, of the character and appearing in any of said windows, coincides with a horizontal pixel location associated with said one first output node

forms, in response to all of the windows, a plurality of output activations for each of the first output nodes, and wherein arrays of the image pixels, each situated in a corresponding one of said windows, are each successively applied to corresponding ones of said plurality of first input nodes of said first neural network in order to generate said output activations, and

means, responsive to plurality of activation outputs that occur for all of the windows, for ascertaining the position, with respect to the field, of the pre-defined element of the character wherein said ascertaining means has;

means for averaging over time the activation output generated by each of said first output nodes and produced for all of said windows to yield an average activation value for each of said first output nodes;

means for filtering the average activation value for each of the first output nodes to yield an index value for each of said first output nodes;

means for thresholding the index values, against a pre-defined threshold value, for said first nodes thereby yielding an activation group formed of thresholded index values; and

means responsive to thresholded index values within the activation group, for determining the position of the pre-defined element of the character.

2. The system in claim 1 wherein the position of the pre-defined element is a center of the character and wherein the array portion is centrally located within each of said windows.

3. The system in claim 2 wherein said extracting means parses a portion of the field containing said character and as defined by the position of the pre-defined element so as to yield a parsed field portion, and comprises means for recognizing the one character in said parsed field portion and for providing an output representing said one character.

4. The system in claim 3 further comprising means, responsive to the image field containing the character, for defining a sub-image containing the character, for normalizing the sub-image to a pre-defined size and for applying the sub-image to the extracting means as the field.

5. The system in claim 3 wherein said recognizing means comprises a character classifier to which each of the pixels in said parsed field portion being routed as input and for providing the output representing the one character.

6. The system in claim 5 wherein said character classifier is a second neural network having a plurality of second input nodes and plurality of output nodes, each of the image pixels in the parsed field portion being routed to a different corresponding one of the second input nodes and each of the second output nodes represents a different one of a plurality of symbols in a character set.

7. The system in claim 6 wherein the second neural network comprises:

an input layer, a plurality of hidden layers and an output layer.

8. The system in claim 7 wherein a first one of the hidden layers in said second neural network is a Gabor layer and each of the remaining hidden layers in the second neural network is a trainable layer.

9. The system in claim 5 wherein said recognizing means further comprises means, responsive to the position of the pre-defined element of the character and a confidence measure provided by the character classifier, for validating a character recognized by said character classifier as a valid character and for providing the valid character as the output representing said one character.

10. The system in claim 2 wherein the averaging means forms a weighted average of said activation outputs generated by each of said first output nodes, wherein each of said activation outputs is weighted by a corresponding one of a first set of weights, each of said first weights being a pre-defined number so as to favor activation outputs occurring in a central section of the array portion of each of said windows.

11. The system in claim 2 wherein said filtering means forms a weighted combination of predefined groups of averaged activation values for successive ones of the first output nodes to yield corresponding index values.

12. The system in claim 2 wherein said position determining means comprises means for providing the position of the pre-defined element of the character as a weighted average of said thresholded index values, wherein each of said thresholded index values is weighted by a corresponding one of a second set of weights, each of said second weights being a number representative of a pixel position associated with a corresponding one of said thresholded index values.

13. The system in claim 2 wherein the first neural network comprises:

an input layer, a plurality of hidden layers and an output layer.

14. The system in claim 13 wherein a first one of the hidden layers in said first neural network is a Gabor layer and each of the remaining hidden layers in the first neural network is a trainable layer.

15. In an optical character recognition system for recognizing a character situated within an image field, wherein the image field contains image pixels and the system detects a position of a pre-defined element of said character with respect to the field, a method comprising the steps of:

successively extracting pre-defined array portions from said field so as no define a plurality of corresponding windows, all of said windows having a pre-defined size, wherein each of the windows has a starting position relative to the field that is offset, along a horizontal field direction, by a pre-defined pixel amount with respect to the starting position of an adjacent one of the windows relative to the field;

through a first neural network, responsive to said extracting step and having plurality of first input nodes and a plurality of first output nodes, wherein each output node is associated with a different horizontal pixel location within an array portion of each windows the steps of;

successively applying arrays of the image pixels, each array being situated in a corresponding one of said windows, to corresponding ones of first input nodes; and

generating an activation output at one of the first output nodes whenever the position of the pre-defined element, of the character and appearing in any of said windows, coincides with a horizontal pixel location associated with said one first output node so as to form, in response to all of the windows, a plurality of output activations for each of the first output nodes; and

ascertaining, in response to plurality of activation outputs that occur for all of the windows, the position, with respect to the field, of the pre-defined element of the character wherein said position ascertaining step further comprises the steps of;

averaging over time, the activation outputs generated by each of said first output nodes and produced for all of said windows to yield an average activation value for each of said first output nodes;

filtering the average activation value associated with each of the first output nodes to yield an index value for each of said first output nodes;

thresholding the index values, against a pre-defined threshold value, for said first nodes thereby yielding an activation group formed of thresholded index values; and

determining, in response to thresholded index values within the activation group, the position of the pre-defined element of the character.

16. The method in claim 15 wherein the position of the pre-defined element is a center of the character and wherein the array portion is centrally located within each of said windows.

17. The method in claim 16 wherein said extracting step comprises the steps of:

parsing a portion of the field containing said character and as defined by the position of the pre-defined element so as to yield a parsed field portion;

recognizing the one character in said parsed field portion; and

providing an output representing said one character.

18. The method in claim 17 further comprising the steps, in response to the image field containing the character, of:

defining a sub-image containing the character;

normalizing the sub-image to a pre-defined size; and

applying the sub-image to the extracting step as the field.

19. The method in claim 17 wherein said recognizing step comprises the steps of:

classifying a character in said parsed field portion being routed as input; and

providing a classification output as the output representing the one character.

20. The method in claim 19 wherein said recognizing step further comprises the steps of:

validating, in response to the position of the pre-defined element of the character and a confidence measure provided by the character classifying step, a character recognized by said character classifier as a valid character; and

providing the valid character as the output representing said one character.

21. The method in claim 16 wherein the averaging step comprises the step of forming a weighted average of said activation outputs generated by each of said first output nodes, wherein each of said activation outputs is weighted by a corresponding one of a first set of weights, each of said first weights being a predefined number so as to favor activation outputs occurring in a central section of the array portion of each of said windows.

22. The method in claim 16 wherein said filtering step comprises the step of forming a weighted combination of pre-defined groups of averaged activation values for successive ones of the first output nodes to yield corresponding index values.

23. The method in claim 16 wherein said position determining step comprises the step of providing the position of the pre-defined element of the character as a weighted average of said thresholded index values, wherein each of said thresholded index values is weighted by a corresponding one of a second set of weights, each of said second weights being a number representative of a pixel position associated with a corresponding one of said thresholded index values.