Pattern recognition neural network

• US 5,440,651 A
• Filed: 04/20/1993
• Issued: 08/08/1995
• Est. Priority Date: 06/12/1991
• Status: Expired due to Fees

• Current Assignee: Microelectronics and Computer Technology Corporation
• Original Assignee: Microelectronics and Computer Technology Corporation
• Inventors: Martin, Gale L.
• Primary Examiner(s): Mancuso, Joseph
• Application Number:
  US08/049,967
• Time in Patent Office: 840 Days
• Field of Search: 382/14, 382/15, 382/9, 382/36, 395/10, 395/20, 395/21, 395/62, 395/77
• 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 30/262 using context analysis, e.g. lexical, syntactic or semantic context

Abstract

A multi-layered pattern recognition neural network that comprises an input layer (28) that is operable to be mapped onto an input space comprising a scan window (12). Two hidden layers (30) and (32) map the input space to an output layer (16). The hidden layers utilize a local receptor field architecture and store representations of objects within the scan window (12) for mapping into one of a plurality of output nodes. Each of the plurality of output nodes and associated representations stored in the hidden layer define an object that is centered within the scan window (12). When centered, the object and its associated representation in the hidden layer result in activation of the associated output node. The output node is only activated when the character is centered in the scan window (12). As the scan window (12) scans a string of text, the output nodes are only activated when the associated character moves within the substantial center of the scan window. The network is trained by backpropagation through various letter string such that the letter by itself within the substantial center of the scan window (12) will be recognized, and also the letter with constraints of additional letters on either side thereof will also be recognized. In addition, the center between characters is recognized when it is disposed substantially in the center of scan window (12), and a space is recognized when it is disposed within the substantial center of the scan window (12).

Claims

1. A multi-layered neural network for pattern recognition, comprising:

an input layer for mapping into a two-dimensional scan window, said scan window defining an input space that can be scanned over objects and which said scan window is sized to encompass more than one object;

an output layer comprised of at least one output node to represent the presence of a desired object substantially centered in said scan window; and

a hidden layer having local receptor fields and interconnected with said input layer and said output layer for mapping said input layer to said output layer, said hidden layer providing a stored representation of the position of said desired object in said scan window relative to the position when said desired object is substantially centered within said scan window such that a determination of the position of said desired object relative to the substantial center of said scan window can be made, wherein said stored representation is trained on said desired object at multiple positions in said scan window at different activation levels for select ones of said positions;

said at least one output node operable to be activated at a predetermined level when an object is disposed at a given position within said scan window, which given position in said scan window and object corresponds to said stored representations, said predetermined level being a function of the relative position of said desired object in said scan window to the substantial center of said scan window.

2. The neural network of claim 1, wherein said output layer comprises a plurality of output nodes corresponding to a plurality of desired objects which exist in a predetermined sequence and can be disposed adjacent each other and overlapping, and wherein said hidden layer contains a stored representation of the relative position of each of said desired objects to the substantial center of said scan window, such that when an object corresponding to any of said stored representations of said desired objects is disposed within said scan window, activation of the associated one of said output nodes results with a level representative of its relative position within said scan window.

3. The neural network of claim 2, wherein an additional one of said output nodes in said output layer is provided representing the presence of the center between two of said desired objects, and said hidden layer contains a stored representation of the relative position of the center between two of said desired objects to the substantial center of said scan window.

4. The neural network of claim 2, wherein said plurality of objects comprises letters of the alphabet.

5. The neural network of claim 2, wherein said plurality of objects comprises numbers from "0" through "9".

6. The neural network of claim 1, and further comprising an additional output node in said output layer for representing the position of the absence of an object within of said scan window relative to the substantial center of said scan window, and said hidden layer contains a stored representation of the position of the absence of a desired object within said scan window relative to the substantial center of said scan window.

7. The neural network of claim 1, wherein said stored representation of the portion of said desired object within said scan window is learned by backpropagation learning techniques.

8. The neural network of claim 1, wherein said hidden layer comprises a first hidden layer having local receptor fields having shared weights and a second hidden layer having local receptor fields and no shared weights.

9. The neural network of claim 1, wherein said desired objects in said scan window are normalized in size to said scan window.

10. A method for recognizing a pattern, comprising:

providing an input layer in a neural network;

mapping the input layer into a two dimensional scan window, the scan window defining an input space that can be scanned over objects and sized to encompass more than one of the objects;

providing an output layer in the neural network having at least one output node having multiple output levels that are operable to represent the position of a desired object within the scan window relative to a predetermined point in the scan window;

providing a hidden layer in the neural network and interconnecting the hidden layer with the input and the output layer;

mapping the input layer to the output layer with the hidden layer, the hidden layer providing a stored representation of the position of the desired object in the in the scan window relative to the predetermined point in the scan window such that a determination of the portion of the desired object relative to the substantial center of the scan window can be made, wherein the stored representation is trained on the desired object at multiple positions in the scan window at different activation levels for select ones of the positions; and

activating the at least one output node in response to an object being disposed in the scan window and substantially corresponding to the stored representation in the hidden layer, the level of activation corresponding g to the position of the object in the scan window.

11. The method of claim 10, wherein the output layer comprises a plurality of output nodes that correspond to a plurality of desired objects that exist in a predetermined sequence and which desired objects can be disposed adjacent to each other and overlapping and comprising:

storing representations of each of the desired objects in the hidden layer that represent the position of the desired objects relative to the predetermined point in the scan window;

each of the representations associated with one of the output nodes; and

activating the associated one of the output nodes when the associated one of the objects or a substantially similar representation thereof is disposed in the scan window, the output level of the activated output node corresponding to the position of the object within the scan position.

12. The method of claim 11 and further comprising:

providing an additional output node in the output layer;

storing in the hidden layer a representation of the position of the center between two of the plurality of desired objects relative to the predetermined point in the scan window; and

activating the additional output node when the center between two objects, corresponding to the stored representation of the position center between two of the desired objects, is disposed in the scan window.

13. The method of claim 11, wherein the plurality of objects comprises letters of the alphabet.

14. The method of claim 10 and further comprising:

providing an additional output node in the output layer;

storing in the hidden layer the representation of the position of the absence of the desired objects relative to the predetermined point in the scan window; and

activating the additional output node when the position of the absence of the desired objects exists within the scan window corresponds to the stored representation of the position of the absence of the desired objects.

15. The method of claim 10 and further comprising training the neural network through backpropagation learning techniques to store in the hidden layer the representation of the position of the desired objects relative to the predetermined point in the scan window.

16. The method of claim 10, wherein the step of providing the hidden layer comprises:

providing a first hidden layer having local receptor fields, having shared weights; and

providing a second hidden layer having local receptor fields and no shared weights.

17. The method of claim 10 and further comprising normalizing objects that are disposed within the scan window to a normalized size.

18. A method for training a neural network, comprising the steps of:

providing a neural network having a input layer and an output layer with stored representations provided in a hidden layer that map the input layer to the output layer, the output layer having a plurality of output nodes;

defining a scan window that is sized to encompass more than one object;

mapping the scan window into the input of the neural network;

passing a desired object through the scan window; and

training the neural network to store a representation of multiple positions of the desired object within the scan window relative to a predetermined point within the scan window such that a given level of activation energy will be associated with at least one of the output nodes when a substantially identical object is disposed at a corresponding position within the scan window that corresponds with the stored representation of different ones of the trained positions of the desired object corresponding to different activation levels for the at least one of the output nodes.

19. The method of claim 18, wherein the step of training comprises training the neural network with backpropagation techniques.

20. The method of claim 18, wherein the step of passing a desired object through the scan window comprises passing a plurality of different desired objects through the the scan window, and the step of training the network comprises training the network to store a representation of the position of each of the different desired objects relative to the predetermined point in the scan window, with an output node associated with each one of the different desired objects, with each output node exhibiting a given level of activation energy for each position of the associated desired object in the scan window when a substantially similar object is disposed at a corresponding position in the scan window that corresponds to stored representation associated with the output node.

21. The method of claim 18, wherein the step of training comprises training the network with a plurality of generalized objects that are substantially similar to the desired object such that the representations stored in the hidden layer of the neural network for a given object is generalized.