WISTI: a simple efficient Textuo-Visual Web Image Retrieval model - Specifications and Benchmarks

Sabrina Tollari (1) , Hervé Glotin (1)

1 LSIS - Laboratoire des Sciences de l'Information et des Systèmes

In this article we specify the textual and visual algorithms that we developed and merged to design a simple efficient Textuo-Visual Web Image Retrieval System that we call WISTI (Web Image Search by Text and Image content). It integrates subband entropy profile visual features and usual mean and color standard deviation, which distributions are depicted. A simple weighted norm fusion is done with tf-idf web page analysis. WISTI is the second best model (af- ter XEROX) according to the official European ImagEVAL Technovision 2006 campaign evaluation. We depict analyses of the fusion behavior of each of the 25 queries on 700 URLs, with Mean Average Precision curves for different text ratio. These precise analyses bring us to propose and discuss on "visualness" versus "textualness" of each query concept.

Mots clés: recherche d'images web ImagEVAL texte

Domaines: Recherche d'information [cs.IR]

Sabrina Tollari, Hervé Glotin. WISTI: a simple efficient Textuo-Visual Web Image Retrieval model - Specifications and Benchmarks. Workshop ImagEVAL 2006, Jul 2007, France. ⟨hal-00199834⟩