Training image classifiers with similarity metrics, linear programming, and minimal supervision

Karl Ni, Ethan Phelps, Katherine L. Bouman, Nadya Bliss

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Image classification is a classical computer vision problem with applications to semantic image annotation, querying, and indexing. Recent and effective generative techniques assume Gaussianity, rely on distance metrics, and estimate distributions, but are unfortunately not convex nor keep computational architecture in mind. We propose image content classification through convex linear programming using similarity metrics rather than commonly-used Mahalanobis distances. The algorithm is solved through a hybrid iterative method that takes advantage of optimization space properties. Our optimization problem uses dot products in the feature space exclusively, and therefore can be extended to non-linear kernel functions in the transductive setting.

Original languageEnglish (US)
Title of host publicationConference Record - Asilomar Conference on Signals, Systems and Computers
Pages1979-1983
Number of pages5
DOIs
StatePublished - 2012
Externally publishedYes
Event46th Asilomar Conference on Signals, Systems and Computers, ASILOMAR 2012 - Pacific Grove, CA, United States
Duration: Nov 4 2012Nov 7 2012

Other

Other46th Asilomar Conference on Signals, Systems and Computers, ASILOMAR 2012
CountryUnited States
CityPacific Grove, CA
Period11/4/1211/7/12

Fingerprint

Linear programming
Classifiers
Image classification
Iterative methods
Computer vision
Semantics

ASJC Scopus subject areas

  • Computer Networks and Communications
  • Signal Processing

Cite this

Ni, K., Phelps, E., Bouman, K. L., & Bliss, N. (2012). Training image classifiers with similarity metrics, linear programming, and minimal supervision. In Conference Record - Asilomar Conference on Signals, Systems and Computers (pp. 1979-1983). [6489386] https://doi.org/10.1109/ACSSC.2012.6489386

Training image classifiers with similarity metrics, linear programming, and minimal supervision. / Ni, Karl; Phelps, Ethan; Bouman, Katherine L.; Bliss, Nadya.

Conference Record - Asilomar Conference on Signals, Systems and Computers. 2012. p. 1979-1983 6489386.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Ni, K, Phelps, E, Bouman, KL & Bliss, N 2012, Training image classifiers with similarity metrics, linear programming, and minimal supervision. in Conference Record - Asilomar Conference on Signals, Systems and Computers., 6489386, pp. 1979-1983, 46th Asilomar Conference on Signals, Systems and Computers, ASILOMAR 2012, Pacific Grove, CA, United States, 11/4/12. https://doi.org/10.1109/ACSSC.2012.6489386
Ni K, Phelps E, Bouman KL, Bliss N. Training image classifiers with similarity metrics, linear programming, and minimal supervision. In Conference Record - Asilomar Conference on Signals, Systems and Computers. 2012. p. 1979-1983. 6489386 https://doi.org/10.1109/ACSSC.2012.6489386
Ni, Karl ; Phelps, Ethan ; Bouman, Katherine L. ; Bliss, Nadya. / Training image classifiers with similarity metrics, linear programming, and minimal supervision. Conference Record - Asilomar Conference on Signals, Systems and Computers. 2012. pp. 1979-1983
@inproceedings{5a5a6874ce3b4f96b17b5e83cef4e638,
title = "Training image classifiers with similarity metrics, linear programming, and minimal supervision",
abstract = "Image classification is a classical computer vision problem with applications to semantic image annotation, querying, and indexing. Recent and effective generative techniques assume Gaussianity, rely on distance metrics, and estimate distributions, but are unfortunately not convex nor keep computational architecture in mind. We propose image content classification through convex linear programming using similarity metrics rather than commonly-used Mahalanobis distances. The algorithm is solved through a hybrid iterative method that takes advantage of optimization space properties. Our optimization problem uses dot products in the feature space exclusively, and therefore can be extended to non-linear kernel functions in the transductive setting.",
author = "Karl Ni and Ethan Phelps and Bouman, {Katherine L.} and Nadya Bliss",
year = "2012",
doi = "10.1109/ACSSC.2012.6489386",
language = "English (US)",
isbn = "9781467350518",
pages = "1979--1983",
booktitle = "Conference Record - Asilomar Conference on Signals, Systems and Computers",

}

TY - GEN

T1 - Training image classifiers with similarity metrics, linear programming, and minimal supervision

AU - Ni, Karl

AU - Phelps, Ethan

AU - Bouman, Katherine L.

AU - Bliss, Nadya

PY - 2012

Y1 - 2012

N2 - Image classification is a classical computer vision problem with applications to semantic image annotation, querying, and indexing. Recent and effective generative techniques assume Gaussianity, rely on distance metrics, and estimate distributions, but are unfortunately not convex nor keep computational architecture in mind. We propose image content classification through convex linear programming using similarity metrics rather than commonly-used Mahalanobis distances. The algorithm is solved through a hybrid iterative method that takes advantage of optimization space properties. Our optimization problem uses dot products in the feature space exclusively, and therefore can be extended to non-linear kernel functions in the transductive setting.

AB - Image classification is a classical computer vision problem with applications to semantic image annotation, querying, and indexing. Recent and effective generative techniques assume Gaussianity, rely on distance metrics, and estimate distributions, but are unfortunately not convex nor keep computational architecture in mind. We propose image content classification through convex linear programming using similarity metrics rather than commonly-used Mahalanobis distances. The algorithm is solved through a hybrid iterative method that takes advantage of optimization space properties. Our optimization problem uses dot products in the feature space exclusively, and therefore can be extended to non-linear kernel functions in the transductive setting.

UR - http://www.scopus.com/inward/record.url?scp=84876270372&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84876270372&partnerID=8YFLogxK

U2 - 10.1109/ACSSC.2012.6489386

DO - 10.1109/ACSSC.2012.6489386

M3 - Conference contribution

AN - SCOPUS:84876270372

SN - 9781467350518

SP - 1979

EP - 1983

BT - Conference Record - Asilomar Conference on Signals, Systems and Computers

ER -