Object recognition system employing a sparse comparison neural network

• US 5,293,456 A
• Filed: 06/28/1991
• Issued: 03/08/1994
• Est. Priority Date: 06/28/1991
• Status: Expired due to Term

• Current Assignee: E. I. du Pont de Nemours and Company (Corteva, Inc.)
• Original Assignee: E. I. du Pont de Nemours and Company (Corteva, Inc.)
• Inventors: Guez, Ygal G., Stafford, Richard G.
• Primary Examiner(s): Fleming, Michael R.
• Assistant Examiner(s): Downs, Robert W.
• Application Number:
  US07/723,348
• Time in Patent Office: 984 Days
• Field of Search: 395/21, 395/22, 395/23, 395/24
• US Class Current: 706/20
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

A neural network for comparing a known input to an unknown input comprises a first layer for receiving a first known input tensor and a first unknown input tensor. A second layer receives the first known and unknown input tensors. The second layer has at least one first trainable weight tensor associated with the first known input tensor and at least one second trainable weight tensor associated with the first unknown input tensor. The second layer includes at least one first processing element for transforming the first known input tensor on the first trainable weight tensor to produce a first known output and at least one second processing element for transforming the first unknown input tensor on the second trainable weight tensor to produce a first unknown output. The first known output comprises a first known output tensor of at least rank zero and has a third trainable weight tensor associated therewith. The first unknown output comprises a first unknown output tensor of at least rank zero and has a fourth trainable weight tensor associated therewith. The first known output tensor and the first unknown tensor are combined to form a second input tensor. A third layer receives the second input tensor. The third layer has at least one fifth trainable weight tensor associated with the second input tensor. The third layer includes at least one third processing element for transforming the second input tensor on the fifth trainable weight tensor, thereby comparing the first known output with the first unknown output and producing a resultant output. The resultant output is indicative of the degree of similarity between the first known input tensor and the first unknown input tensor.

Claims

1. An object recognition system for recognizing a known object in an image, comprising:

(a) means for generating a first known tensor representation of an image of at least one known object, the known object being represented by a first known tensor and for generating a first unknown tensor representation of an image of at least one candidate, unknown object, the unknown object being represented by a first unknown tensor;

(b) a comparison neural network comprising a plurality of parallel processing networks, each parallel processing network including;

(i) a first layer including a first input node for receiving the first known tensor and a second input node for receiving the first unknown tensor,

(ii) a second layer for receiving the first known and unknown tensors, the second layer having at least one first trainable weight tensor associated with the first known tensor and at least one second trainable weight tensor associated with the first unknown tensor, the second layer including first processing means for transforming the first known tensor on the first trainable weight tensor to produce a first known output, the first known output comprising a first known output tensor of at least rank zero having a third trainable weight tensor associated therewith, the second layer further including second processing means for transforming the first unknown tensor on the second trainable weight tensor to produce a first unknown output, the first unknown output comprising a first unknown output tensor of at least rank zero having a fourth trainable weight tensor associated therewith, the first known output tensor and the first unknown output tensor being combined to form a second input tensor having a fifth trainable weight tensor associated therewith,

(iii) a third layer for receiving the second input tensor, the third layer including third processing means for transforming the second input tensor on the fifth trainable weight tensor, thereby comparing the first known output with the first unknown output and producing a resultant output, wherein the resultant output is indicative of the degree of similarity between the first known tensor and the first unknown tensor;

(c) a selection criterion module for receiving and comparing the resultant output of each parallel processing network, the selection criterion module producing an outcome based on a predetermined selection criterion, wherein the outcome is indicative of the closest degree of similarity between the known object and the candidate, unknown object; and

(d) a designating layer for designating the candidate object having the closest degree of similarity to the known object.

2. The neural network as claimed in claim 1, wherein the first known tensor for each parallel processing network is different.

3. The neural network as claimed in claim 1, wherein the first unknown tensor for each parallel processing network is the same.

4. A method of recognizing a known object in an image, comprising the steps of:

(a) generating a first known tensor representation of an image of a known object for each of a plurality of parallel processing networks, the known object being represented by a first known tensor;

(b) generating a first unknown tensor representation of an image of at least one candidate, unknown object for each of the parallel processing networks, the unknown object being represented by a first unknown tensor;

(c) transforming the first known tensor on a first trainable weight tensor associated therewith to produce a first known output for each parallel processing network, the first known output comprising a first known output tensor of at least rank zero;

(d) transforming the first unknown tensor on a second trainable weight tensor to produce a first unknown output for each parallel processing network, the first unknown output comprising a first unknown output tensor of at least rank zero;

(e) combining the first known output tensor and the first unknown output tensor to form a second input tensor for each parallel processing network, the second input tensor having a third trainable weight tensor associated therewith;

(f) transforming the second input tensor on the third trainable weight tensor for each parallel processing network, thereby comparing the first known output with the first unknown output;

(g) producing a resultant output for each parallel processing network, wherein the resultant output is indicative of the degree of similarity between the first known tensor and the first unknown tensor;

(h) comparing the resultant output of each parallel processing network;

(i) producing an outcome indicative of the closest degree of similarity between the known object and the candidate, unknown object; and

(j) designating the candidate object having the closest degree of similarity to the known object.