Predicting performance metrics for algorithms

Patent metadata

Field	Value
Publication number	US-9830704-B1
Application number	US-201615284420-A
Country	US
Kind code	B1
Filing date	Oct 3, 2016
Priority date	Feb 23, 2015
Publication date	Nov 28, 2017
Grant date	Nov 28, 2017

Abstract

Where a plurality of machine learning algorithms is available to process information or data in the furtherance of a task, one of the algorithms may be identified as particularly well-suited or appropriate based on attributes of the information or data. Such attributes of the imaging data may be determined by any means, and a prediction as to the performance (e.g., one or more metrics) or success of each of the algorithms may be made. One of the algorithms may ultimately be selected based on such predictions, as well as the computing resources that are available for executing the algorithms, and any other relevant constraints.

First claim

What is claimed is:

1. A vehicle comprising:
an imaging device having an imaging sensor;
a position sensor;
at least one memory device; and
at least one computer processor,
wherein the at least one computer processor is configured to at least:
capture imaging data using the imaging device, wherein the imaging data comprises a plurality of image frames, and wherein at least one of the plurality of image frames depicts a plurality of alphanumeric characters;
determine at least one spectral property of the at least one of the plurality of image frames;
determine a position of the vehicle by the position sensor when the at least one of the plurality of image frames was captured;
identify information regarding a plurality of recognition algorithms using the at least one computer processor;
predict at least one performance metric of each of the plurality of recognition algorithms for recognizing the plurality of alphanumeric characters depicted in the imaging data based at least in part on the at least one spectral property and the position of the vehicle;
select one of the plurality of recognition algorithms based at least in part on the at least one predicted performance metric; and
execute the selected one of the plurality of recognition algorithms for recognizing the plurality of alphanumeric characters depicted in the imaging data using the at least one computer processor.

2. The vehicle of claim 1 , wherein the at least one computer processor is further configured to at least:
derive a spectral signature for the imaging data based at least in part on the at least one spectral property, wherein the spectral signature comprises a vector representative of the at least one spectral property, and
wherein the at least one performance metric of each of the plurality of recognition algorithms is predicted based at least in part on the spectral signature.

3. The vehicle of claim 1 , wherein the at least one performance metric is one of:
an amount of memory required to execute the at least one of the plurality of recognition algorithms using the at least one computer processor;
a level of accuracy or precision of output of the at least one of the plurality of recognition algorithms;
a rate of success of the at least one of the plurality of recognition algorithms; or
a time required to execute the at least one of the plurality of recognition algorithms using the at least one computer processor.

4. The vehicle of claim 1 , wherein the at least one spectral property comprises at least one of:
a value of an intensity of at least one pixel in the at least one of the plurality of image frames;
a representation of a texture in the at least one of the plurality of image frames;
a representation of an edge component in the at least one of the plurality of image frames;
an indicator of pixel homogeneity or pixel heterogeneity in the at least one of the plurality of image frames; or
a label associated with at least a subset of the at least one of the plurality of image frames.

5. A computer-implemented method comprising:
capturing imaging data by at least one imaging device;
determining at least one spectral property of the imaging data by at least one computer processor;
predicting, for each of a plurality of algorithms, at least one performance metric for performing a predetermined task based at least in part on the imaging data by the at least one computer processor; and
selecting one of the plurality of algorithms for performing the predetermined task on the imaging data based at least in part on the at least one predicted performance metric by the at least one computer processor.

6. The computer-implemented method of claim 5 , further comprising:
executing the selected one of the plurality of algorithms by the at least one computer processor using the imaging data as an input; and
storing an output from the execution of the selected one of the plurality of algorithms in at least one data store.

7. The computer-implemented method of claim 6 , wherein the predetermined task is one of:
geolocating the imaging device at a time when the imaging data was captured based at least in part on contents of the imaging data;
interpreting a plurality of characters expressed in the imaging data;
recognizing an object depicted in the imaging data; or
interpreting a bar code associated with the object depicted in the imaging data.

8. The computer-implemented method of claim 5 , wherein determining the at least one spectral property of the imaging data further comprises:
providing at least some of the imaging data as an input to at least one filtering algorithm;
processing the at least some of the imaging data according to the at least one filtering algorithm using the at least one computer processor; and
determining the at least one spectral property of the imaging data based at least in part on an output of the at least one filtering algorithm using the at least one computer processor.

9. The computer-implemented method of claim 8 , wherein the at least one filtering algorithm is one of a fast Fourier transform or a discrete Fourier transform.

10. The computer-implemented method of claim 5 , further comprising:
identifying at least one background feature expressed in the imaging data; and
determining the at least one spectral property of the imaging data based at least in part on the at least one background feature,
wherein the at least one spectral property is at least one of:
at least one color of a portion of the imaging data corresponding to the at least one background feature;
at least one dimension of the at least one background feature;
at least one representation of at least one horizontal texture or edge component of the at least one background feature;
at least one representation of at least one vertical texture or edge component of the at least one background feature;
at least one representation of at least one directional texture or edge component of the at least one background feature;
an indicator of chromaticity of the portion of the imaging data corresponding to the at least one background feature;
an indicator of color saturation of the portion of the imaging data corresponding to the at least one background feature;
an indicator of at least one pixel feature of the portion of the imaging data corresponding to the at least one background feature in at least one scale or level of resolution;
an indicator of pixel heterogeneity in the portion of the imaging data corresponding to the at least one background feature; or
an indicator of pixel homogeneity in the portion of the imaging data corresponding to the at least one background feature.

11. The computer-implemented method of claim 5 , wherein determining the at least one spectral property of the imaging data comprises:
determining a plurality of spectral properties of the imaging data by the at least one computer processor, and
wherein the method further comprises:
generating a spectral signature based at least in part on at least some of the plurality of spectral properties of the imaging data,
wherein the at least one performance metric for performing the predetermined task is predicted for each of the plurality of algorithms based at least in part on the spectral signature.

12. The computer-implemented method of claim 11 , wherein generating the spectral signature comprises:
obtaining a discrete Fourier transform based at least in part on at least some of the imaging data,
wherein the spectral signature comprises a plurality of pixel intensity frequency representations determined based at least in part on the discrete Fourier transform.

Assignees

• Amazon Tech Inc

Inventors

• Mishra Pragyana K

Classifications

• H04N5/77 Primary

  between a recording apparatus and a television camera · CPC title

• G06V10/87

  using selection of the recognition techniques, e.g. of a classifier in a multiple classifier system · CPC title

• H04N7/181

  for receiving images from a plurality of remote sources · CPC title

• G06F18/285

  Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system · CPC title

• G06K9/00201

  Physics · mapped topic