Automatic image annotation with long distance spatial-context

Abstract

Because of high computational complexity, a long distance spatial-context based automatic image annotation is hard to achieve. Some state of art approaches in image processing, such as 2D-HMM, only considering short distance spatial-context (two neighbors) to reduce the computational complexity. However, these approaches cannot describe long distance semantic spatial-context in image. Therefore, in this paper, we propose a two-step Long Distance Spatial-context Model (LDSM) to solve that problem. First, because of high computational complexity in 2D spatial-context, we transform a 2D spatial-context into a 1D sequence-context. Second, we use conditional random fields to model the 1D sequence-context. Our experiments show that LDSM models the semantic relation between annotated object and background, and experiment results outperform the classical automatic image annotation approach (SVM).