Saliency oriented object image re-ranking

Abstract

For image retrieval, most users hope to retrieve a certain salient object instead of an obscure pattern in an image. This paper presents a novel object re-ranking algorithm based on visual saliency, which is employed to detect salient object regions in an image. The re-ranking is carried out with a SVM classifier on the salient regions to assure that the images ranked on the top of the list exhibit salient object pictures. To speed up the classifier training, a small code book (1K) is chosen. To improve the re-ranking efficacy, we employ the posterior probability of the salient region to adjust re-ranking, and derive an approximate formula of the posterior probability of the salient region. The formula is based on a hierarchical model, containing spatial information to compensate the feature disorder of the model of bags of visual words (BoVW). The posterior probability is calculated offline, so the online efficiency of re-ranking is high. Experiments demonstrate that our algorithm significantly improves online efficiency and saliency while possesses high accuracy of image retrieval.