Image Labeling and Classification by Semantic Tag Analysis
Kirbac, Ugur (2013)
2013
Master's Degree Programme in Information Technology
Tieto- ja sähkötekniikan tiedekunta - Faculty of Computing and Electrical Engineering
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2013-05-08
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201305231168
https://urn.fi/URN:NBN:fi:tty-201305231168
Tiivistelmä
Image classification and retrieval plays a significant role in dealing with large multimedia data on the Internet. Social networks, image sharing websites and mobile application require categorizing multimedia items for more efficient search and storage. Therefore, image classification and retrieval methods gained a great importance for researchers and companies.
Image classification can be performed in a supervised and semi-supervised manner and in order to categorize an unknown image, a statistical model created using pre-labeled samples is fed with the numerical representation of the visual features of images.
A supervised approach requires a set of labeled data to create a statistical model, and subsequently classify an unlabeled test set. However, labeling images manually requires a great deal of time and effort. Therefore, a major research activity has gravitated to wards finding efficient methods to reduce the time and effort for image labeling. Most images on social websites have associated tags that somewhat describe their content. These tags may provide significant content descriptors if a semantic bridge can be established between image content and tags. In this thesis, we focus on cases where accurate class labels are scarce or even absent while some associated tags are only present. The goal is to analyze and utilize available tags to categorize database images to form a training dataset over which a dedicated classifier is trained and then used for image classification. Our framework contains a semantic text analysis tool based on WordNet to measure the semantic relatedness between the associated image tags and predefined class labels, and a novel method for labeling the corresponding images. The classifier is trained using only low-level visual image features. The experimental results using 7 classes from MirFlickr dataset demonstrate that semantically analyzing tags attached to images significantly improves the image classification accuracy by providing additional training data.
Image classification can be performed in a supervised and semi-supervised manner and in order to categorize an unknown image, a statistical model created using pre-labeled samples is fed with the numerical representation of the visual features of images.
A supervised approach requires a set of labeled data to create a statistical model, and subsequently classify an unlabeled test set. However, labeling images manually requires a great deal of time and effort. Therefore, a major research activity has gravitated to wards finding efficient methods to reduce the time and effort for image labeling. Most images on social websites have associated tags that somewhat describe their content. These tags may provide significant content descriptors if a semantic bridge can be established between image content and tags. In this thesis, we focus on cases where accurate class labels are scarce or even absent while some associated tags are only present. The goal is to analyze and utilize available tags to categorize database images to form a training dataset over which a dedicated classifier is trained and then used for image classification. Our framework contains a semantic text analysis tool based on WordNet to measure the semantic relatedness between the associated image tags and predefined class labels, and a novel method for labeling the corresponding images. The classifier is trained using only low-level visual image features. The experimental results using 7 classes from MirFlickr dataset demonstrate that semantically analyzing tags attached to images significantly improves the image classification accuracy by providing additional training data.