Diffusion-based Image Saliency Detection

Visual saliency detection is an important technology in computer vision.Human visual system has the ability to select important information for further processing.The mechanism like a filter can reserve interesting information,and make the content close to image itself.The goal of saliency detection is to extract a region containing the salient object.Saliency detection was widely used in many computer vision applications including image retrieval,segmentation,and object recognition etc.Nowadays,the diffusion-based salient object detection is an important method among a lot of saliency detection method,Given an initial set of saliency seeds,different diffusion mode is used to propagate saliency information.The key is how to choose these seeds and propagate their saliency.Given the inconspicuous difference of color and contrast between the background and the salient object,the erroneous seeds are identified.Finally,these wrong seeds affect the quality of saliency map.In this paper,the initial saliency map is obtained and threshold it to identify very salient pixels.Saliency information is then propagated using effective diffusion process and an optimized saliency map is finally obtained.The main contributions of our paper are:(1)Estimating an initial saliency map using existing salient region detection techiques after 6 approaches are evaluated on a dataset.(2)The initial saliency map is divided different patches,and each patch is thresholded to obtain a map.The map is to identify very salient pixels that definitely belong to salient regions and backgrounds.(3)The new propagating algorithm which combine the regularized random walks and manifold ranking can generate pixel-wised saliency maps that reflect the details of the input images.(4)We have evaluated our approach on two public datasets(MSRA-1000 and MSRA-5000)and clothing dataset was created by our research group.Experiment results show that our algorithm is effective outperformed,existing saliency detection methods,yielding higher precision and better recall rates.