Computationally efficient SVM multi-class image recognition with confidence measures

Lázaro Makili, Jesús Vega, Sebastián Dormido-Canto, Ignacio Pastor, Andrea Murari

Research output: Contribution to journalArticle

14 Citations (Scopus)


Typically, machine learning methods produce non-qualified estimates, i.e. the accuracy and reliability of the predictions are not provided. Transductive predictors are very recent classifiers able to provide, simultaneously with the prediction, a couple of values (confidence and credibility) to reflect the quality of the prediction. Usually, a drawback of the transductive techniques for huge datasets and large dimensionality is the high computational time. To overcome this issue, a more efficient classifier has been used in a multi-class image classification problem in the TJ-II stellarator database. It is based on the creation of a hash function to generate several "one versus the rest" classifiers for every class. By using Support Vector Machines as the underlying classifier, a comparison between the pure transductive approach and the new method has been performed. In both cases, the success rates are high and the computation time with the new method is up to 0.4 times the old one. © 2011 Elsevier B.V. All Rights Reserved.
Original languageEnglish
Pages (from-to)1213 - 1216
Number of pages4
JournalFusion Engineering and Design
Issue number6-8
Publication statusPublished - Oct 2011
Externally publishedYes


All Science Journal Classification (ASJC) codes

  • Civil and Structural Engineering
  • Materials Science(all)
  • Nuclear Energy and Engineering
  • Mechanical Engineering

Cite this