Using manifold structure for automatic image annotation by fusion of multiple feature spaces

Abstract

Automatic image annotation has been an active research topic in recent years. Low level features like as color, texture, shape as well as object spatial relations are extracted to represent images in general. These syntaxes are further used to retrieve images from large image data sets.  However, the similarity of images could not be found correctly by similarity measures such as Euclidean distance in many situations. On the other hand, graph models have been shown powerful in solving many machine learning problems in recent years. In this paper, we propose a graph-based learning approach, named Conceptual Manifold Structure (CMS), based on transition from conceptual to observation space. In the proposed method, a graph including both the trained and tested samples is constructed by fusion of multiple feature spaces. Conceptual transition in graph structure is found by altering the edge values in an innovative manner. This is caused to learn the manifold structure where the samples dissimilarity is closer to the conceptual distance. Furthermore, the continuity between the instances of a semantic in the conceptual space is kept in feature space. Keeping the continuity in manifold structure is the main idea to decrease the semantic gap in this study. The experiments on different image data sets indicated that the geometrical distances between the samples on the manifold space are closer to their conceptual distance. The proposed method has been compared to other well-known approaches. The results confirmed the effectiveness and validity of the proposed method.