Robust clustering for social annotated images

Clustering plays an important role in large-scale image data management. The supervised clustering methods can discover the meaningful data groups according to the image labels. However, social annotated labels usually contain lots of noises like semantic ambiguity and redundancy, and thus the accuracy of clustering cannot be guaranteed. This paper firstly analyzes the characteristics of social annotated labels: label co-occurrence, ambiguity and redundancy, and then proposes a new image clustering method using semantic labels and their co-occurrence statistics and designs the similarity metric for social annotated images. Moreover, this paper presents a dynamic and efficient thresholding scheme for adaptively terminate the spectral clustering process. Finally, a social annotated image data set is constructed for algorithm evaluation. In our experiments, we compared our method with classic ones, and the results show that our method has better robustness and efficiency on social annotated images in terms of both accuracy and balance.