Back to search

Athens University Economics and Business (AUEB)

Multimedia object classification and retrieval

Abstract

dc:description

In this thesis the problem of multimedia object classification and retrieval is studied. The first part examines the application of Latent Semantic Analysis to large scale image retrieval. An effective and efficient approach of applying LSA is presented skipping the SVD solution of the feature matrix and overcoming in this way the deficiencies of the method with large scale datasets. The study demonstrates that early fusion of several composite descriptors with visual words increase retrieval effectiveness. It also combines well with late fusion for mixed (textual and visual) ad hoc and modality classification. The results reported are comparable to state of the art algorithms. In addition, we propose an LSA based kernel function which effectively integrates low-level visual features with higher semantic ones into a common latent space that correlates multi-modal features (features from different sources) in the same latent space. The proposed approach also effectively incorporates classification with image retrieval in an early fusion manner. Experimental results show that it outperforms the latent indexing generated by singular value decomposition.In terms of image representation, a generalization of the Bag-of-Colors (BoC) model is proposed. The new algorithm, referred to as QBoC, is based on a quad tree decomposition of images and manages to efficiently encode spatial information in the final image representation. Combined with the Bag-of-Visual-Words (BoVW) model it is used to efficiently identify the modality detection of images in the medical domain. Finally, a novel approach for combining LSA with Deep Convolutional Neural Networks (CNN) for image classification is presented. In particular, the potential of Neural Networks when images are represented by compact descriptors is investigated. To this end, an optimized latent semantic space is constructed that captures the affinity of images to each modality using a pre-trained network. The images are represented by a Weighted Latent Semantic Tensor in a lower space and they are used to train a deep CNN that makes the final classification. Experimental results demonstrate the effectiveness and the efficiency of our framework in terms of classification accuracy, achieving comparable results to current state-of-the-art approaches.

Degree

thesis:*
Grantor dc:publisher
Athens University Economics and Business (AUEB)
Year dc:date
2018

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Stathopoulos, Spyridon
  • Σταθόπουλος, Σπυρίδωνας

Subjects

dc:subject × 13

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/44612

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Stathopoulos, Spyridon; Σταθόπουλος, Σπυρίδωνας. Multimedia object classification and retrieval. Athens University Economics and Business (AUEB), 2018. http://hdl.handle.net/10442/hedi/44612