Research on middle-semantic manifold object annotation

Abstract

A novel bionic, middle-semantic object annotation framework is presented in this paper. Moreover, we build the model based on the perception as defined by the human visual system. At first, the super-pixel is used to represent the images, and conditional random field could label each of the super-pixels, which means annotating the different classes of objects. In next step, on the basis of the previous result, image pyramid is used to represent the image, and get the sub-region of some objects of the same class. After extracting descriptor to represent the patches, all the patches are projected to a manifold, which could annotate the different views of objects from the same class. Experiments show that the bionic, middle-semantic object annotation framework could obtain superior results with respect to accuracy, and it could verify the correctness of WordNet indirectly. © Springer-Verlag Berlin Heidelberg 2014.