Similarity ranking technique exploiting the structure of similarity relationships
- PDF / 1,887,759 Bytes
- 18 Pages / 439.37 x 666.142 pts Page_size
- 5 Downloads / 230 Views
Similarity ranking technique exploiting the structure of similarity relationships Guang-Ho Cha1 Received: 29 April 2020 / Accepted: 17 October 2020 © Springer-Verlag GmbH Austria, part of Springer Nature 2020
Abstract This paper proposes a similarity ranking technique that exploits the entire network structure of similarity relationships for multimedia, particularly image, databases. The main problem in the similarity ranking on multimedia is the meaning gap between the characteristics automatically computed from the multimedia dataset and the interpretation by human from the multimedia itself. In fact, the similarity semantics usually lies on high level human interpretation and automatically computed low level multimedia properties may not reflect it. This paper assumes that the meaning of the multimedia is affected by the context or similarity relationships in a dataset and therefore, we propose the ranking technique to catch the semantics from a large multimedia dataset. This similarity ranking technique based on the context or similarity relationships yields better experimental results than the conventional similarity ranking techniques. Keywords Multimedia retrieval · Image retrieval · Semantic learning · Similarity search · Similarity relationship · Similarity ranking · Context relationship Mathematics Subject Classification 68P10 · 68P20
1 Introduction and background The similarity query or the k-nearest neighbor (k-NN) query is the typical query type in multimedia information retrieval. One of key components of constructing effective content-based multimedia retrieval (CBMR) mechanism is how to effectively match the human perception with the multimedia features computed by machine. In CBMR, media are commonly represented by features in high dimensional data spaces and users generally choose query images to find images similar to those. Differently from the
B 1
Guang-Ho Cha [email protected] Department of Computer Science and Engineering, Seoul National University of Science and Technology, Seoul 01811, Republic of Korea
123
G.-H. Cha Fig. 1 Human perception-based similarity search: a a query digit image, b, c user-retrieved images
Fig. 2 Euclidean distance-based similarity search: a a query digit image, b, c machine-retrieved images
conventional information retrieval, the feature vectors utilized in CBMR usually have visual characters, and thus they are not directly connected to the perceptual recognition perceived by users as contextual relationships. The similarity degree between two media objects in conventional CBMR systems is usually calculated by Euclidean distance, more generally, Minkowski metric or L p -norm. But, users have sometimes experienced mismatches among their queries and result images returned from CBMR systems. Actually, the similarity notion commonly lies on high level semantics by human, and the system-computed properties utilized in the similarity computation do not usually include the user’s recognition. The semantic gap is the main challenge to overcome in CBMR research
Data Loading...
