Evaluating web image context extraction

Abstract

Images on theWeb appear with other textual contents—referred to as Web Image Context—providing valuable information to the image semantics. Unfortunately, HTML documents are usually cluttered with multiple different contents to different topics and therefore the right image context needs to be precisely determined in order to deliver high quality descriptions. Several methods that automatically determine and extract the Web image context from Web documents have been applied in different applications over the years. However, in these applications context extraction is only a preprocessing step and therefore the quality of the extraction task has rather been evaluated on its own. To sum up, there is hardly information about which extraction method to choose in order to get best results. Keeping this necessity in mind, an evaluation framework that objectively measures and compares the quality of different Web Image Context Extraction (WICE) algorithms will be the main subject in this book chapter. The main parts of the framework are a large ground truth dataset consisting of diverse Web documents from real Web servers and objective quality measures tailored to fit the special characteristics of the image context extraction task. In order to demonstrate the capabilities of the framework, common extraction methods from the literature are implemented and integrated into the framework. Finally, the evaluation results are summarized and discussed.