Automatic image annotation is a highly challenging problem.Base on cross-media relevance model, this paper presents an approach to annotate images by fusing image classification.In the annotation refinement process,the frequent patterns mining algorithm is used to refine the annotation results. Firstly, image features are extracted to generate visual words so as to describe each image. Secondly, the similarity relationship of images is generated by K-means clustering and the classification information is generated by support vector machines. Then, by knowing the relationships between the semantic labels and the images, we can use statistical methods to calculate the probability of each semantic label. The candidate semantic labels are determined by fusing the classification information of the image as weight into the probability. Finally,based on the probability of candidate label words, an improved frequent patterns mining algorithm is used to mine the text relevance degree. The candidate annotation wordset is processed by equal-frequency discretization to obtain the final annotation results. Experiments on Corel image set achieve a better annotation result.

 

 ,