Výběr reprezentativních obrazových prototypů

Abstract

The diversity of image retrieval results is an important feature that allows users to explore different aspects of the queried object. However, most of the works in this area are focused more on relevance rather than diversity. In this thesis, we are ameliorating this situation by proposing a new method for retrieving a diverse set of landmark images, which is based on recent advances in image retrieval area. The proposed approach consists of three phases. On the first one, irrelevant images are deleted from the input set using two detector networks. Then, the clustering phase follows, where landmark images are divided into groups by visual similarity based on distances between state-of-the-art image descriptors. Finally, from each cluster, a single representative located in the densest area of the cluster is chosen. For each phase, several alternative options are proposed, and the best combination is determined on the MediaEval dataset. Conducted experiments show that the proposed approach is superior to the current state-of-the-art, both in terms of diversity of the retrieved set and in terms of relevance.The diversity of image retrieval results is an important feature that allows users to explore different aspects of the queried object. However, most of the works in this area are focused more on relevance rather than diversity. In this thesis, we are ameliorating this situation by proposing a new method for retrieving a diverse set of landmark images, which is based on recent advances in image retrieval area. The proposed approach consists of three phases. On the first one, irrelevant images are deleted from the input set using two detector networks. Then, the clustering phase follows, where landmark images are divided into groups by visual similarity based on distances between state-of-the-art image descriptors. Finally, from each cluster, a single representative located in the densest area of the cluster is chosen. For each phase, several alternative options are proposed, and the best combination is determined on the MediaEval dataset. Conducted experiments show that the proposed approach is superior to the current state-of-the-art, both in terms of diversity of the retrieved set and in terms of relevance.