Web image indexing using WICE and a learning-free language model
Date Issued
September 2016
Author(s)
DOI
https://doi.org/10.1007/978-3-319-44944-9_12
Abstract
With the advent of Web 2.0 and the rapidly increasing popularity of online social networks that make extended use of visual information, like Facebook and Instagram, web image indexing regained great attention among the researchers in the areas of image indexing and information retrieval. Web image indexing is traditionally approached, by commercial search engines, using text-based information such as image file names, anchor text, web-page keywords and, of course, surrounding text. In the latter case, for effective indexing, two requirements should be met: Correct identification of the related text, known as image context, and extraction of the right terms from this text. Usually, researchers working in the field of web image indexing consider that once the image context is identified extraction of indexing terms is trivial. However, we have shown in our previous work that this is not the rule of thumb.
In this paper we get advantage of Web Image Context Extraction (WICE) using visual web-page parsing and specific distance metrics and following this we locate key terms within this text to index the image using language models. In this way, the proposed method is totally learning free, i.e., no corpus need to be collected to train the keyword extraction component, while the identified indexing terms are more descriptive for the image since they are extracted from a portion of web-page’s text. This deviates from the traditional web image indexing approach in which keywords are extracted from all text in the web-page. The evaluation, performed on a dataset of 978 manually annotated web images taken from 243 web pages, shows the effectiveness of the proposed approach both in image context extraction and indexing.
In this paper we get advantage of Web Image Context Extraction (WICE) using visual web-page parsing and specific distance metrics and following this we locate key terms within this text to index the image using language models. In this way, the proposed method is totally learning free, i.e., no corpus need to be collected to train the keyword extraction component, while the identified indexing terms are more descriptive for the image since they are extracted from a portion of web-page’s text. This deviates from the traditional web image indexing approach in which keywords are extracted from all text in the web-page. The evaluation, performed on a dataset of 978 manually annotated web images taken from 243 web pages, shows the effectiveness of the proposed approach both in image context extraction and indexing.