Image region label refinement using spatial position relation graph

Abstract

With the exponential growth of massive image data, automatic image annotation is becoming more important in image management and retrieval. Traditional image region annotation methods, through machine learning and low-level visual features, typically yield incorrect annotation results owing to the influence of the Semantic Gap. We herein propose a novel label refinement method for improving the image region annotation results. A spatial position relation graph with co-occurrence relations and spatial position relations among labels is proposed to analyze the latent semantic correlations among image region labels. Moreover, an incremental iterative random-walking algorithm is proposed to reconstruct the region relation graph for detecting non-dependable regions whose labels do not fit the semantic context of an image. Subsequently, a graph matching algorithm with semantic correlation and spatial relation analysis is proposed for non-dependable region label completion. Experiments on Corel5K demonstrate that our proposed spatial-position-relation-graph- based label refinement method can achieve good performance for image region label refinement.