A Relevance Model for Web Image Search
Cheng Thao, Ethan V. Munson · 2003
This article describes the construction of a relevance model for Web image search. Using custom image retrieval software, 24 textual queries were used to retrieve over 5800 images. Each image’s relevance to its query was evaluated by human raters. The Web documents containing these im-ages were analyzed for the presence of text matching the query in each of 53 HTML features. Finally, logistic regres-sion was used to construct the relevance model that best predicted the human ratings from the presence of matching text in HTML features. The resulting relevance model has a precision of over 65 % when applied to our entire sample. It uses a total of thirteen HTML features with image filename and document title being the most important. A number of methodologic issues are discussed and suggestions for fu-ture research are made. 1