Neuromorphic parallel processor

• US 8,401,297 B1
• Filed: 06/28/2012
• Issued: 03/19/2013
• Est. Priority Date: 06/28/2011
• Status: Active Grant

• Current Assignee: Ami Research & Development, LLC
• Original Assignee: Ami Research & Development, LLC
• Inventors: Apostolos, John T., Feng, Judy, Mouyos, William, McMahon, Benjamin
• Primary Examiner(s): LE, BRIAN Q
• Application Number:
  US13/536,494
• Time in Patent Office: 264 Days
• Field of Search: 382/115, 382/159, 382/160, 382/165, 382/170, 382/181, 382/209, 382/217, 382/218, 382/224, 382/225, 382/226, 382/227, 382/228, 382/280, 382/199, 382/219, 382/220, 382/210, 382/278, 706/15, 706/19, 706/20, 706/22, 706/41, 700/47, 700/50
• US Class Current: 382/181
• CPC Class Codes: G06F 18/256 of results relating to different input data, e.g. multimodal recognition
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06V 10/431 Frequency domain transformation; Autocorrelation
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 10/811 the classifiers operating on different input data, e.g. multi-modal recognition
• CPC Class Codes: G06V 10/955 using specific electronic processors

Abstract

A neuromorphic parallel image processing approach that has five (5) functional layers. The first performs a frequency domain transform on the image data generating multiple scales and feature based representations which are independent of orientation. The second layer is populated with feature based representations. The third layer, an object class recognizer layer, are fused using a neuromorphic parallel processor. Fusion of multimodal data can achieve high confidence, biometric recognition.

Claims

1. An apparatus comprising a pattern recognizer further comprising:

a first functional layer providing an image processor;

a second functional layer extracting a feature based representation of one or more objects of interest;

a third functional layer providing object class recognition;

wherein image representations are maintained as complex-data values, the first functional layer comprises a complex gradient determination, the second functional layer comprises a complex conjugate chirp fourier transform, and the third functional layer is implemented using a complex valued neuron emulator based on a mixed mode signal processor comprising a complex multiplier, up converter, and dispersive delay line; and

wherein the third functional layer further provides scale and orientation independence by sweeping over expected scales and orientations using a combination of scale zoom, angular shift and oriented templates, to generate a scatter plot of correlation peaks for each different object in the data base, and to determine a final correlation peak location by calculating a centroid of the correlation peaks, with the final peak value determined by adding all contributions from the correlation peaks.

2. The apparatus of claim 1 wherein the third functional layer further fuses results of matching three or more biometric parameters.

3. The apparatus of claim 2 wherein object class recognition further comprises:

storing a set of templates for each of the biometric parameters for each of the members of a population;

obtaining biometric parameters from a subject;

convolving the subject biometric parameters with each of the set of templates to determine a signal to noise ratio (SNR) associated with the response of each template to the respective biometric; and

ranking the results of convolving to produce a rank order list.

4. The apparatus of claim 3 wherein a verification function comprises:

obtaining personal identification information from a subject person; and

providing the templates as representative cross sections of the population.

5. The apparatus of claim 3 wherein SNR is further determined by

squaring all the template voltages in the rank order list to obtain a template power rank order list,

calculating a 2 sigma point on the power rank order list,

calculating an average of the sum of all the template powers greater than the 2 sigma point,

dividing each template power in the rank order list by the average from to obtain the template SNR rank order list.

6. The apparatus of claim 5 wherein SNR is further determined by:

for each biometric,

converting the template SNR rank order list to a new power rank order list by raising 10 to the SNR/10 power for each template,

determining a fusion template power rank order list by adding the new powers from all the biometrics for each respective template; and

determining a fusion template SNR rank order list using the above SNR algorithm.

7. The apparatus of claim 1 wherein a scale/zoom operation is performed on image data such that scaled versions of a template representation need be stored in the second functional layer, and where the templates are correlated against one or more possible zoom states.

8. The apparatus of claim 1 wherein the third functional layer further:

applies a sliding angular shift function in the plane of the image, such that no angular shifted versions of any representation need to be stored in the second layer, and such that the templates are correlated against each angular shift state.

9. An apparatus comprising a pattern recognizer further comprising:

a first functional layer providing an image processor;

a second functional layer extracting a feature based representation of one or more objects of interest;

a third functional layer providing object class recognition;

wherein;

image representations are maintained as complex-data values, the first functional layer comprises a complex gradient determination, the second functional layer comprises a complex conjugate chirp fourier transform, and the third functional layer is implemented using a complex valued neuron emulator based on a mixed mode signal processor comprising a complex multiplier, up converter, and dispersive delay line

wherein object class recognition further comprises

storing a set of templates for each of the biometric parameters for each of the members of a population;

obtaining biometric parameters from a subject;

convolving the subject biometric parameters with each of the set of templates to determine a signal to noise ratio (SNR) associated with the response of each template to the respective biometric; and

ranking the results of convolving to produce a rank order list; and

wherein a verification function comprises;

obtaining personal identification information from a subject person; and

providing the templates as representative cross sections of the population; and

wherein SNR is further determined by

squaring all the template voltages in the rank order list to obtain a template power rank order list,

calculating a 2 sigma point on the power rank order list,

calculating an average of the sum of all the template powers greater than the 2 sigma point,

dividing each template power in the rank order list by the average from to obtain the template SNR rank order list; and

wherein, for each biometric SNR is further determined by;

converting the template SNR rank order list to a new power rank order list by raising 10 to the SNR/10 power for each template,

determining a fusion template power rank order list by adding the new powers from all the biometrics for each respective template; and

determining a fusion template SNR rank order list using the above SNR algorithm.