Composition Based Semantic Scene Retrieval for Ancient Murals

Retrieval of similar scenes in ancient murals research is an important but time-consuming job for researchers. However, content-based image retrieval (CBIR) systems cannot fully deal with such issues since they lack of the abilities to handle complex semantic and image composition queries. In this paper, we introduce a new semantic scene-retrieval approach for ancient murals. Our method can retrieve related scenes according to both their content elements and their composition through a two-phase procedure. Then, retrieved scenes are ranked according to composition-based criterion that incorporates the relevance of semantic content and visual structures with scene compactness ratio. Hence, the sorted results are tailored to the real intent of query. The experiments demonstrate the efficiency and effectiveness of our approach to reduce the semantic gap of visual information retrieval. Furthermore, the retrieval results for Dunhuang murals suggest the potential applications for general paintings retrieval and personalized publishing.