Spatial image processor

• US 6,801,655 B2
• Filed: 05/10/2001
• Issued: 10/05/2004
• Est. Priority Date: 05/10/2001
• Status: Expired due to Fees

• Current Assignee: The United States of America As Represented By The Secretary of Agriculture
• Original Assignee: the united states of america as represented by the secretary of the navy
• Inventors: Woodall, Roger L.
• Primary Examiner(s): Mehta, Bhavesh M.
• Assistant Examiner(s): CHAWAN, SHEELA C
• Application Number:
  US09/853,932
• Publication Number:
  US 20020168100A1
• Time in Patent Office: 1,244 Days
• Field of Search: 382/155-161, 382/305, 704/232, 704/259, 701/40, 701/59, 706/15-44
• US Class Current: 382/156
• CPC Class Codes: G06T 7/0002 Inspection of images, e.g. flaw detection
• CPC Class Codes: G06T 7/20 Analysis of motion motion estimation for coding, decoding, compressing or decompressing digital video signals H04N19/43, H04N19/51
• CPC Class Codes: G06V 30/194 References adjustable by an adaptive method, e.g. learning

Abstract

A spatial image processor neural network for processing image data to discriminate between first and second spatial configurations of component objects includes a photo transducer input array for converting an input image to pixel data and sending the data to a localized gain network (LGN) module, a parallel memory processor and neuron array for receiving the pixel data and processing the pixel data into component recognition vectors and chaotic oscillators for receiving the recognition vectors and sending feedback data to the LGN module as attention activations. The network further includes a temporal spatial retina for receiving both the pixel data and temporal feedback activations and generating temporal spatial vectors, which are processed by a temporal parallel processor into temporal component recognition vectors. A spatial recognition vector array receives the temporal component recognition vectors and forms an object representation of the first configuration of component objects.

Claims

1. A spatial image processor neural network for processing an image so as to recognize a whole object from rearranged and partial configurations of component objects, each component object forming a part of the whole object, said network comprising:

a photo transducer input array for converting the image to pixel data and sending signals indicative of the pixel data;

a localized gain network module for receiving the signals indicative of the pixel data, increasing the gain of said pixel data in response to attention activations received by said module and outputting localized pixel data;

a parallel memory processor and neuron array for receiving the localized pixel data from said localized gain network module, for processing the localized pixel data into component recognition vectors and peripheral vision object activations;

a component recognition vectors assembly for receiving said component recognition vectors, for receiving associative connections feedback, for generating temporal activations, and for sending feedback data to said parallel memory processor and neuron array;

a temporal spatial retina for receiving the localized pixel data from said localized gain network module and said temporal activations from said component recognition vectors assembly, for generating temporal spatial vectors and for sending signals indicative of said temporal activations;

a temporal parallel memory processor for receiving said temporal spatial vectors from said temporal spatial retina and for generating temporal component recognition vectors; and

a temporal, spatial and object recognition vector array for receiving said temporal component recognition vectors from said temporal parallel memory processor, for forming attention activations to the whole object, for sending prototype object activations of a class of objects to which the whole object belongs, and for sending said associative connections feedback to said component recognition vectors assembly for the component objects recognized as forming a part of the whole object.

2. The network in accordance with claim 1, wherein, in said component recognition vectors assembly, each component recognition vector is operable to activate a chaotic oscillator, each of said chaotic oscillators being different, each to represent one of the component objects.

3. The network in accordance with claim 2, wherein each of said chaotic oscillators includes a matched filter and an initial condition which provides an input of synchronizing values and a matched output from each of said oscillators, which matched output provides said feedback data flowing from said component recognition vectors assembly to said parallel memory processor and neuron array.

4. The network in accordance with claim 1, wherein said component recognition vectors assembly is adapted for sequential generation of said temporal activations.

5. The network in accordance with claim 4, wherein said temporal, spatial and object recognition vector array further comprises:

a delay vector array receiving the temporal component recognition vectors and aligning a period of temporal activations of the temporal component recognition vectors to generate synchronized vector outputs;

a plurality of synchronized processing units, each receiving a synchronized vector output from an associated one of the delay vectors in the delay vector array and providing recognition connections;

a plurality of super object neurons, one of the super object neuron being activated when at least one recognition connection is made to the one super object neuron from one of the synchronized processing units, each activated super object neuron providing an object output;

a threshold unit receiving each object output and generating the attention activations to the whole object and the associated connections feedback; and

a prototype classification neuron receiving and summing the synchronized vector outputs to generate the prototype object activations.

6. The network in accordance with claim 5, wherein each said synchronized processing unit further comprises:

a threshold and hold multi-king-of-the-mountain (THMKOM) unit receiving, gating and holding the synchronized vector output from the associated delay vector;

a vector decoupler to decouple the gated and held synchronized vector;

a neural director receiving the decoupled synchronized vector to increase a resolution of the decoupled synchronized vector; and

a positional king-of-the-mountain (PKOM) unit receiving the increased resolution vector to generate a highest element of the increased resolution vector representing the recognition connection of the whole object.

7. The network in accordance with claim 6, wherein said temporal, spatial and object recognition vector array further comprises a threshold and hold module applying a threshold to the synchronized vector outputs of the delay vector array.

8. The network in accordance with claim 1, wherein said localized gain network module includes a vector normalizer for adjusting the received image pixel data into a normalized vector of said image pixel data.

9. The network in accordance with claim 8, wherein said parallel memory processor and neuron array further comprises an image position and size invariant retina generating feature vectors from which are developed said component recognition vectors and said peripheral vision object activations.

10. The network in accordance with claim 9, wherein said parallel memory processor and neuron array further comprises:

multi-layer neural directors receiving said feature vectors; and

feedback and threshold positional king of mountain (FTPKOM) units, each unit receiving a neural director output from an associated one of the multi-layer neural directors, and each unit generating a FTPKOM output.

11. The network in accordance with claim 10, wherein parallel memory processor and neuron array further comprises:

SUM neurons, each SUM neuron receiving the FTPKOM output from an associated one of the FTPKOM units, each SUM neuron being operative to provide a SUM output;

a component memory vector space receiving said SUM outputs and forming connection sets based on activations in the memory vector space elicited by said SUM outputs; and

a neuron array, which, together with the connection sets, forms the component recognition vectors.

12. The network in accordance with claim 9, wherein the image position and size invariant retina further comprises:

a window difference neural network (WDN);

a vector decoupler (VD); and

a gradient window, the WDN forming differences from a reference pixel to all other pixels in the gradient window, said VD being operative to decouple and disperse said differences to provide said feature vectors.

13. The network in accordance with claim 1, wherein said temporal spatial retina is provided with a temporal retina and a spatial retina cooperating to provide said temporal spatial vectors directed to said temporal parallel memory processor.

14. The network in accordance with claim 1, wherein said temporal parallel memory processor further comprises a temporal image position and size invariant retina receiving said temporal spatial vectors and generating temporal feature vectors.

15. The network in accordance with claim 14, wherein said temporal parallel memory processor further comprises:

temporal multi-layer neural directors receiving said feature vectors and generating temporal neural director outputs; and

threshold positional king-of-the-mountain (TPKOM) units, each receiving the neural director output from an associated neural director and each generating a TPKOM output.

16. The network in accordance with claim 15, wherein said temporal parallel memory processor further comprises:

temporal SUM neurons, each receiving the TPKOM output from an associated TPKOM unit and providing a temporal SUM output;

a temporal component memory vector space receiving said temporal SUM outputs and forming temporal connection sets based on activations in the temporal component memory vector space elicited by said SUM outputs; and

a temporal neuron array, which, together with the temporal connection sets, forms said temporal component recognition vectors.

17. The network in accordance with claim 1, wherein:

the parallel memory processor and neuron array and the temporal parallel memory processor are combined to form a single memory processor;

a gradient retina applies a gradient to the localized pixel data prior to the localized data being received by the temporal spatial retina, the gradient for each pixel corresponding to a distance between the pixel and a fovea center of the image; and

each component recognition vector is operable to activate a component temporal generator of an array of component temporal generators in the component recognition vectors assembly, the activated component temporal generator applying a sequential pulse corresponding to said temporal activations.