Fuzzy based contextual cueing for region level annotation

This paper investigates the challenging issue of assigning given image-level annotations to precise regions on natural images. We propose a novel label to region assignment (LRA) technique called Fuzzy-based Contextual-cueing Label Propagation (FCLP) with four parts: First, an image is over-segmented into a set of atomic patches and the local visual information of color features and texture features are extracted. Second, fuzzy representation and fuzzy reasoning are used to model contextual cueing information, especially for the imprecise position information and ambiguous spatial topological relationships. Third, labels are propagated inter images in visual space and intra images in contextual cueing space. Finally, the fuzzy C-means clustering based on K-nearest neighbor (KNN-FCM) is utilized to segment the images into semantic regions and associate with corresponding annotations. Experiments on the public datasets demonstrate the effectiveness of the proposed technique.