Naslov
Extracting keywords from images: bag-of-visual- words enriched with graph techniques

Autori
Madjarov, Gjorgji ; Martinčić-Ipšić, Sanda

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, neobjavljeni rad, znanstveni

Izvornik
Challenge Track, International Keystone Conference / Martinčić-Ipšić, Sanda ; Horvat, Marko - Coimbra, 2015

Skup
Challenge Track, International Keystone Conference 2015

Mjesto i datum
Coimbra, Portugal, 08.09.2015. - 09.09.2015

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
keyword extraction; images; bag-of-visual-words; graph techniques

Sažetak
Keywords have become primary means for searching information in documents, images and videos on the WWW. Automatic keyword extraction establishes foundation for various natural language and multimedia processing applications: information retrieval, automatic indexing and classification of a collection of documents, automatic summarization, high-level semantic description, etc. The task of keyword extraction is to automatically identify a set of terms that best describe the document. State-of-the-art keyword extraction approaches are based on statistical methods which require learning from hand-annotated data sets. Lately, the focus of research has shifted toward unsupervised methods, mainly network or graph enabled keyword extraction has attracted researcher’s attention. In a network (graph) based keyword extraction the source (document, text, specific data etc.) is transformed into network in a way: words (or units) are nodes of the network and their relations are represented with links. This way, both the statistical properties (frequencies) as well as the structure of source are represented by unique formal representation, hence complex network. Graph formalism beside text can model many different data sources biological, ecological, social relations, transporting infrastructure, etc. We propose extending graph representation model for bag- of-visual words (BoVW) used in image retrieval for extracting the most representative visual parts of image using graph-enabled keyword extraction principles. State-of-the-art systems for image retrieval use BoVW representation of images. In BoVW models, a vocabulary (or codebook) of visual words is obtained by clustering local image descriptors extracted from images. An image is then represented as a BoVW, which is a sparse vector of occurrence counts of the visual words in the vocabulary.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Informacijske i komunikacijske znanosti

POVEZANOST RADA