A Study On Semantic-based Thangka Image Annotation And Retrieval

This paper combined with characters of Thangka image itself, and chose headdress and gesture of Thangka image as subjects to start developing the research of semantic-based Thangka image annotation and retrieval, on the basis of review the SBIR technologies. Firstly, we proposed a headdress feature expression approach, which mainly used contour feature and color feature. The average of spectral frequency represents contour feature. The color moment of internal contour represents color feature. Meanwhile, we can automatically annotate headdress concepts through computing variable quantity of spectral frequency and analyzing color distribution of internal contour. Secondly, this paper presented a gesture feature expression method including five fingers direction, palm direction, relate position of gestures etc. Because of the difficult features for segmenting gestures of Thangka image. Then, the concepts of unknown gestures can be predicted with nearest neighbor method on the basis of training annotated gesture set. Thirdly, we applied Delphi method and fussy statistic to construct the C-K relation net, which can descript subordination between semantic keywords and concepts better. Finally, semantic-based Thangka image retrieval was roughly classified into three layers in this paper: (1) it is similarities of visual features that receive most attention on visual layer; (2) it towards coherence of key area's concept on concept layer; (3) high level semantic retrieval place more weight on similarities of concepts. The experiment show that different retrieval layers can meet different users'demand and outcome is satisfactory.