Systems and methods for optimized computer vision using deep neural networks and Litpschitz analysis

Patent metadata

Field	Value
Publication number	US-10839253-B2
Application number	US-201916443118-A
Country	US
Kind code	B2
Filing date	Jun 17, 2019
Priority date	Jun 15, 2018
Publication date	Nov 17, 2020
Grant date	Nov 17, 2020

Abstract

Computer vision systems and methods for optimized computer vision using deep neural networks and Lipschitz analysis are provided. The system receives signals or data related to visual imagery, such as data from a camera, and feed-forwards the signals/data through the multiple layers of a convolutional neural network (CNN). At one or more layers of the CNN, the system determines at least one Bessel bound of that layer. The system then determines a Lipschitz bound based on the one or more Bessel bounds. The system then applies the Lipschitz bound to the signals. Once the Lipschitz bound is applied, the system can feed-forward the signals to other processes of the layer or to a further layer.

First claim

What is claimed is:

1. A computer vision system, comprising:
at least one computer system; and
computer vision system code executed by the at least one computer system, the computer vision system code causing the computer system to:
receive, at a first layer of a convolution neural network executing on the computer system, a first input node wherein the first input node comprises a first input signal;
generate a first output signal by passing the first input signal through a first filter;
determine at least one first Bessel bound for the first layer;
determine a first Lipschitz bound based on the at least one first Bessel bound; and
apply the first Lipschitz bound to the first output signal.

2. The computer vision system of claim 1 , wherein the computer vision system code further causes the computer system to:
transform the first output signal into an output node;
feed-forward the output node to a second layer of the convolution neural network, wherein the output node becomes a second input node, the second input node comprising a second input signal;
generate a second output signal by passing the second input signal through a second filter;
determine at least one second Bessel bound for the second layer;
determine a second Lipschitz bound based on the at least one first Bessel bound and the at least one second Bessel bound; and
apply the second Lipschitz bound to the output signal.

3. The computer vision system of claim 1 , wherein the first filter can be one of a convolution filter, a merge filter or a pooling filter.

4. The computer vision system of claim 1 , wherein the at least one first Bessel bound is determined based on whether the first layer comprises a merge filter.

5. The computer vision system of claim 1 , wherein the at least one first Bessel bound comprises three Bessel bounds.

6. The computer vision system of claim 5 , wherein each of the three Bessel bounds are determined by a different formula.

7. The computer vision system of claim 6 , wherein the first Lipschitz bound is determined by using a value from at least one of the three Bessel bounds.

8. A method, comprising,
receiving, at a first layer of a convolution neural network executing on a computer system, a first input node wherein the first input node comprises a first input signal;
generating a first output signal by passing the first input signal through a first filter;
determining at least one first Bessel bound for the first layer;
determining a first Lipschitz bound based on the at least one first Bessel bound; and
applying the first Lipschitz bound to the first output signal.

9. The method of claim 8 , further comprising:
transforming the first output signal into an output node;
feed-forwarding the output node to a second layer of the convolution neural network, wherein the output node becomes a second input node, the second input node comprising a second input signal;
generating a second output signal by passing the second input signal through a second filter;
determining at least one second Bessel bound for the second layer;
determining a second Lipschitz bound based on the at least one first Bessel bound and the at least one second Bessel bound; and
applying the second Lipschitz bound to the output signal.

10. The method of claim 8 , wherein the first filter can be one of a convolution filter, a merge filter or a pooling filter.

11. The method of claim 8 , wherein the at least one first Bessel bound is determined based on whether the first layer comprises a merge filter.

12. The method of claim 8 , wherein the at least one first Bessel bound comprises three Bessel bounds.

13. The method of claim 12 , wherein each of the three Bessel bounds are determined by a different formula.

14. The method of claim 13 , wherein the first Lipschitz bound is determined by using a value from at least one of the three Bessel bounds.

15. A non-transitory, computer-readable medium having computer readable instructions stored thereon which, when executed by a processor, cause the processor to perform the steps comprising,
receiving, at a first layer of a convolution neural network, a first input node wherein the first input node comprises a first input signal;
generating a first output signal by passing the first input signal through a first filter;
determining at least one first Bessel bound for the first layer;
determining a first Lipschitz bound based on the at least one first Bessel bound; and
applying the first Lipschitz bound to the first output signal.

16. The computer-readable medium of claim 15 , further comprising computer-readable instructions for causing the processor to perform the steps of:
transforming the first output signal into an output node;
feed-forward the output node to a second layer of the convolution neural network, wherein the output node becomes a second input node, the second input node comprising a second input signal;
generating a second output signal by passing the second input signal through a second filter;
determining at least one second Bessel bound for the second layer;
determining a second Lipschitz bound based on the at least one first Bessel bound and the at least one second Bessel bound; and
applying the second Lipschitz bound to the output signal.

17. The computer-readable medium of claim 15 , wherein the first filter can be one of a convolution filter, a merge filter or a pooling filter.

18. The computer-readable medium of claim 15 , wherein the at least one first Bessel bound is determined based on whether the first layer comprises a merge filter.

19. The computer-readable medium of claim 15 , wherein:
the at least one first Bessel bound comprises three Bessel bounds; and
each of the three Bessel bounds are determined by a different formula.

20. The computer-readable medium of claim 19 , wherein the first Lipschitz bound is determined by using a value from at least one of the three Bessel bounds.

Assignees

• Insurance Services Office Inc

Inventors

• Balan Radu
• Singh Maneesh Kumar
• Zou Dongmian

Classifications

• G06V10/82

  using neural networks · CPC title

• G06V10/454

  Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN] · CPC title

• G06F18/21

  Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06N3/045

  Combinations of networks · CPC title