Abstract: An image semantic relevance annotation method is proposed,which aims at tagging large-scale image collections in real environment.First,in specific feature space,each training image is encoded as a sparse linear combination of other training images by sparse representation.Then a multiple label learning method based on maximum a posteriori principle is utilized to generate the relevance of tags and images in specific feature space.Finally,an optimal threshold set for each tag and corresponding feature is estimated,and the tag relevance of diverse features can be combined in the manner of unsupervised combination.The experiments on the Internet image set show the superior performance of the proposed framework.