Salient Object Detection Based On Prior Integration And Manifold Ranking

Salient object detection has been applied to numerous vision problems, including content based image retargeting, object classification and recognition, image figure/ground segmentation, image retrieval, to name a few. Therefore, recent years have witnessed more interest in salient object detection. The main task of salient object detection is to accurately detect where the salient object should be, and generate a saliency map where the intensity of the pixel indicates the likelihood of that pixel belonging to the salient object.Due to the absence of high level knowledge, the bottom-up salient object detection is a highly ill-posed problem. Appropriate prior knowledge exploitation is helpful for ill-posed problems. Thus, this paper proposes a novel prior integration based salient object detection model, which is referred as PI model. Previous saliency detection approaches are almost based on contrast prior. However, contrast prior based methods still have some certain limitations, e.g., the object region cannot be uniformly highlighted. Recently, many works incorporate the center prior into the saliency model. However, the center prior is sensitive to the location of the object and becomes invalid when the objects are placed far off the image center. Considering this issue, this paper applies a rough object detection method to estimate the center of the salient object and assumes that the object appears near this new center, which makes the proposed center prior more robust and effective. The contrast prior and the proposed center prior are combined to compute an initial saliency map. To further improve the initial results, this paper incorporates the smoothness prior, which has been widely used in image segmentation models, into the saliency detection. This paper encourages the similar neighboring pixels or superpixels to sharing similar saliency values though minimizing a continuous pairwise saliency energy function on graph, i.e., the smoothness prior, which can uniformly highlight the object region.In this paper, saliency detection is formulated as a manifold ranking problem and a novel saliency detection model is proposed. The manifold structure of the image data is first represented with a close-loop k-regular graph model. Saliency detection is then carried out in a two-stage scheme to extract background regions and foreground salient objects efficiently. Most existing methods measure the foreground saliency directly by considering the object cues, such as contrast priors, whereas a few methods focus on segmenting out background regions and thereby salient objects. Unlike the existing methods, this paper considers both the background and foreground cues in a different way. The first stage extracts background regions through considering background cues, which can provide sufficient foreground cues for the second stage. The second stage detects the object region directly by exploiting the object cues and the foreground information getting from the first stage.Experiment results on the MSRA-1000and MSRA databases demonstrate the proposed two methods perform well when against the state-of-the-art methods.