Image retrieval is one of the hottest fields of computer vision and pattern recognition. In recent years, many researchers addressed the challenging problem of interpreting the semantics of images. This paper presented a novel approach based on relation net (concept and semantic keyword relation net) for high level semantic retrieval of Thangka image. Here, we use Delphi method and fuzzy statistic to construct the relation net, which can descript the membership degree between semantic keywords and concepts better. Finally, this paper proposed CSM (concept similar measurement) algorithm to compute the similarity during the concepts on the basis of relation net. The experiments show that the proposed approach can retrieval Thangka image by the similar concepts well and effectively.