A fast visual word frequency - inverse image frequency for detector of rare concepts

  • Authors:

  • Emilie Dumont;Hervé Glotin;Sébastien Paris;Zhong-Qiu Zhao

  • Affiliations:

  • Sciences and Information Lab., LSIS, UMR, CNRS, France and University of Sud, Toulon-Var, France;Sciences and Information Lab., LSIS, UMR, CNRS, France and University of Sud, Toulon-Var, France;Sciences and Information Lab., LSIS, UMR, CNRS, France;College of Computer Science and Information Engineering, Hefei University of Technology, China

  • Venue:
  • CLEF'09 Proceedings of the 10th international conference on Cross-language evaluation forum: multimedia experiments
  • Year:
  • 2009

Quantified Score

Hi-index 0.00

Visualization

Abstract

In this paper we propose an original image retrieval model inspired from the vector space information retrieval model. We build for different features and different scales a visual concept dictionary composed by visual words intended to represent a semantic concept, and then we represent an image by the frequency of the visual words within the image. Then the image similarity is computed as in the textual domain where a textual document is represented by a vector in which each component is the frequency of occurrence of a specific textual word in that document. We then adapt the common text-based paradigm by using the TF-IDF weighting scheme to construct a WF-IIF weighting scheme in our Multi-Scale Visual Dictionary (MSVD) vector space model. The experiments are conducted on the 2009 Visual Concept Detection ImageCLEF Campaign. We compare WF-IIF to usual direct Support-Vector Machine (SVM) algorithm. We demonstrate that SVM and WFIIF are in average over all the concept giving the same Area Under the Curve (AUC). We then discuss the fusion process that should enhance the whole system, and of some particular properties of MSVD, that shall be less dependant of the training set size of each concept than the SVM.