Research On Salient Region Detection Algorithm

Salient region detection aims at finding the most distinctive object regions in an image.When we look at an image,distinctive objects can instantly attract our attention.We can get prior target regions through salient region detection,so as to allocate the computing resources in a more reasonable way and reduce the computing consumption when we process the information in an image.Therefore,detecting salient object region from an image has high application value.Generally,saliency detection methods can be categorized as either top-down or bottom-up approaches.Top-down methods are task-driven and require supervised learning with high-level information.This method has a complex cross-disciplinary problem.Because it may be need to combine the knowledge of neurosciences,biology and other related fields.In contrast,bottom-up methods are data-driven and usually exploit low-level cues,such as color contrast and spatial layout features et al.,to obtain the salient object region.The operation of bottom-up methods is simple and fast.Recent research has shown that this method is quite successful and it has become a useful tool in many application scenarios,including image segmentation,object recognition,visual tracking,and so on.In this paper,we focus on the bottom-up salient region detection.We exploit the lowlevel cues such as color contrast and spatial layout feature to detect the salient object regions.And,we proposed two kinds of saliency detection algorithms in total.The one is KLMS,an abbreviation for “salient region detection via KL divergence and multi-scale merging”,the another is SRM,an abbreviation for “salient region detection via multiscale region merging”.Experimental results on some large benchmark datasets demonstrate the proposed two algorithms outperform some state-of-the-art methods,and have higher saliency detection precision and recall rates.The proposed two algorithms can also produce smooth saliency maps.The major research work of the paper includes:1.In the algorithm of KLMS: We segment the original image into super-pixels via SLICO algorithm,then the color values of all pixels in each super-pixel are clustered in terms of their discriminative power to get the statistical probability distribution of the cluster labels for each super-pixel.We measure the color dissimilarity values between each superpixel pairs with the harmonic-mean of KL Divergence of their probability distributions,so as to solve the limitation of leveraging average color difference value.We calculate multi-scale saliency maps according to boundary connectivity and region contrast based on the corresponding undirected close-loop connected graph,in which nodes are the super-pixels and the adjacent regions are expanded reasonably,and edges are weighted with the harmonic-mean of KL Divergence.Then,we integrate all the saliency maps to get the final one,so as to reduce the misjudgment probability of single-scale saliency detection.2.In the algorithm of SRM: Based on the region adjacent graph,we obtain multiscale region merging images via adjacent merging and global merging operation,which can help us make a better hierarchical representation of the original image.We set the Coherence Matrix based on multiscale background probability values of super-pixel region cells,which can help improve the optimized effect of the Cellular Automata.