Saliency Detection Based On Boundary Prior And Contrast Optimization

Saliency detection is a process which focuses on finding salient object or critical region in an given image.With the advent of big data era,image,as a carrier of information,has become an important way to deliver messages.In computer vision area,saliency detection algorithm was widely used in image segmentation,object finding and tracking,image enhancement and so on.Therefore,research on saliency detection and its related applications is of great significance,especially in massive image information processing.Saliency detection algorithms can be roughly divided into top-down approach and bottomup approach.In recent years,many bottom-up methods combined background prior with low cues of image itself and outperformed biological stimulus models.Unfortunately,existing boundary prior based saliency detection algorithms have some problems in selecting reasonable saliency prior section,which would lead to inaccuracy in calculating foreground sections,thus have a negative effect on final result.Aiming at this problem,In this paper,we propose a new salient object detection algorithm,which based on contrast optimized manifold ranking.To take advantage of boundary information of images to find background prior,a method was designed that using saliency expectation,local contrast and global contrast to measure the quality of the background saliency prior and then fused them by weighted sum rather than simple multiplication,so as to acquire accurate saliency estimation.When extracting salient section from prior,the threshold policy was optimized to get more reasonable foreground section.By using graph based manifold ranking,we obtain the final saliency map,which will be more accurate than previous work.According to experiment,the proposed method have a better performance than other congener algorithms.It highlights the salient object in an image with less noise and produces saliency map which was well suited to human visual perception.And it costs less time than deep learning approaches.