Image Partition And Location Based Saliency Detection

Salient object detection has become a hot topic in the fields of computer vision.For lack of high-level semantic information,bottom-up saliency detection methods rely on priors for visual saliency of foreground objects and backgrounds,including contrast,background and center priors.Contrast prior claims that ”salient object presents high contrast within a certain context”.Methods based on contrast prior calculate contrast in the entire image or in a local neighborhoods.Those methods have difficulties in highlighting the whole objects uniformly.Background prior assumes that ”image boundary shows similar properties with backgrounds”.Methods based on background prior use similarity between image elements and boundaries to exclude backgrounds.These methods have great advantages with methods based on contrast prior,but are biased in detecing images with uniform backgrounds.Center prior suggests that salient object is close to image center,which is usually a supplement to other methods.Considering the above three priors,from the point of view of background prior,we firstly split the source image into three parts: ROI(region of interest),image border and intermediate region.ROI and image border are seperately treated as pseudo-foreground and pseudobackground regions to get the likelihoods of patches belonging to foregrounds and backgrounds.Then,as for center prior,we define a location measure to calculate the distance between image center and patches.Finally,foreground likelihood and location measure are regarded as a foreground measure,background likelihood are regraded as a background measure,we use energy optimization function to combine the two measures with patch smoothness to get the saliency model.Experiments demonstrate that the saliency model proposed by our paper can highlight the salient object uniformly and exclude most backgrounds,which outperform most state-of-the-art methods.