Locally connected neural network with improved feature vector

• US 6,757,666 B1
• Filed: 04/13/2000
• Issued: 06/29/2004
• Est. Priority Date: 04/13/1999
• Status: Expired due to Term

• Current Assignee: California Institute of Technology
• Original Assignee: California Institute of Technology
• Inventors: Thomas, Tyson
• Primary Examiner(s): STARKS, WILBERT L
• Application Number:
  US09/549,216
• Time in Patent Office: 1,538 Days
• Field of Search: 706/15, 706/16, 342/159, 378/62, 382/100, 382/157, 382/291
• US Class Current: 706/15
• CPC Class Codes: G06V 10/42 Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation

Abstract

A pattern recognizer which uses neuromorphs with a fixed amount of energy that is distributed among the elements. The distribution of the energy is used to form a histogram which is used as a feature vector.

Claims

1. A method of pattern recognition, comprising:

obtaining an image to be pattern recognized;

forming a feature vector representing said image, said feature vector being indicative of a histogram of information, where each element of the histogram represents a number of parts of the image which have attained a specified energy level; and

applying pixels of the input image to neuromorphs which respond to a feature of the input image, each said neuromorph attaining an energy level based on said features of the image.

2. A method as in claim 1, wherein there is a fixed amount of energy for the entire image, and said neuromorphs apportion said fixed amount of energy.

3. A method as in claim 2, further comprising allowing the network of neuromorphs to settle once an input image is received, and characterizing each of the neuromorphs according to an energy activation level thereof.

4. A method as in claim 3, wherein said specified energy levels are activation levels, and further comprising sensing activation levels of each of the neuromorphs, classifying the activation levels, and forming said histogram based on a number of neuromorphs in a specified level.

5. A method as in claim 4, further comprising assigning a fixed energy amount to the network, and allowing said neuromorphs to distribute said fixed energy amount to form said activation levels.

6. A method as in claim 5, further comprising using the feature vector in a classifier for final recognition of a pattern represented by the image.

7. A method as in claim 1, wherein each said neuromorph forms an input to other neuromorphs and receives outputs from other neuromorphs.

8. A method as in claim 1, wherein said classifying comprises providing a plurality of elements and interconnecting said elements.

9. A method as in claim 1, further comprising weighting the elements.

10. A method of pattern recognition, comprising:

obtaining an image to be pattern recognized; and

forming a feature vector representing said image, said feature vector being indicative of a histogram of information, where each element of the histogram represents a number of parts of the image which have attained a specified energy level;

detecting network activation according to ${\mathrm{\alpha <br><br>}}_{\mathrm{ij}}<br><br>\left(n+1\right)=\frac{I_{\mathrm{ij}}+\underset{\mathrm{kl}\in <br><br>N_r<br><br>\left(\mathrm{ij}\right)}{\overset{\text{\hspace{1em}<br><br>}}{\sum <br><br>}}<br><br>\left[w_{\mathrm{ij}<br><br>\text{\hspace{1em}<br><br>};<br><br>\mathrm{kl}}·<br><br>{\mathrm{\alpha <br><br>}}_{\mathrm{kl}}<br><br>\left(n\right)\right]}{\underset{\mathrm{ij}}{\overset{\text{\hspace{1em}<br><br>}}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>\left(I_{\mathrm{ij}}+\underset{\mathrm{kl}\in <br><br>N_r<br><br>\left(\mathrm{ij}\right)}{\overset{\text{\hspace{1em}<br><br>}}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>\left[w_{\mathrm{ij}<br><br>\text{\hspace{1em}<br><br>};<br><br>\mathrm{kl}}·<br><br>{\mathrm{\alpha <br><br>}}_{\mathrm{kl}}<br><br>\left(n\right)\right]\right)}·<br><br>E$

where,

klε

N_r(ij) are the coordinates kl of a point that falls within a radius r of the neighborhood of neuromorph ij;

α

_kl(n) is the current activation level of neuromorph kl in N_r(ij)

W<sub>ij;

kl</sub> is the weight of the synaptic or linking connection between neuromorph ij and neuromorph kl;

I_ij is the input pattern pixel value at location ij;

E is the global network energy constant;

n is the iteration number.

11. A method of pattern recognition, comprising:

obtaining an image to be pattern recognized;

forming a feature vector representing said image, said feature vector being indicative of a histogram of information, where each element of the histogram represents a number of parts of the image which have attained a specified energy level; and

wherein there is a fixed amount of energy for the entire image, and said neuromorphs apportion said fixed amount of energy; and

adjusting a number of bins in a histogram, which number of bins represents the number of locations where elements representing specified energy amounts can be stored based on an application.

12. A method as in claim 11, wherein said application is fingerprint recognition.

13. A recognition system, comprising:

an image sensor element; and

a pattern recognizing element outputting a feature vector, said feature vector having characteristics which are rotation invariant, said classifying element operating using a fixed amount of energy, and allocating the fixed amount of energy among multiple elements of the classifying circuit depending a feature of the image;

wherein the feature vector represents a histogram indicating how many pixels have specified energy amounts;

wherein the histogram includes a number of neuromorphs that are in each range of activation energy, and said feature vector is indicative of the histogram.

14. A system as in claim 13, wherein the histogram is also substantially invariant to translation.

15. A recognition system, comprising:

an image sensor element;

a pattern recognizing element outputting a feature vector, said feature vector having characteristics which are rotation invariant, said classifying element operating using a fixed amount of energy, and allocating the fixed amount of energy among multiple elements of the classifying circuit depending a feature of the image; and

further comprising providing a fixed current source for energy.

16. A method of recognizing a pattern, comprising:

linking between a plurality of elements, each of which obtains information indicating a pixel of the pattern; and

inhibiting positive feedback in the linking.

17. A method as in claim 16, wherein said inhibiting comprises limiting a total amount of energy available to all of the plurality of elements.

18. A method as in claim 17, further comprising forming a feature vector based on a distribution of the energy.

19. A method as in claim 18, wherein said feature vector is representative of a number of elements in each bin of a histogram of energy distribution.

20. A method as in claim 19, further comprising only using bins that include activations among above a specified threshold.