Saliency Detection Via Surroundedness And Markov Model

Salient object detection is a long-term study of the problem in the field of computer vision,and attract more and more researchers' attention in recent years.The goal of salient object detection is to detect an area of image that attracts human visual attention accurately and efficiently.Efficient and accurate salient object detection has been applied to image processing applications in the fields of Internet and artificial intelligence,such as image retrieval,image compression,image segmentation and content-aware image editing.In this thesis,we propose a method to detect image saliency by using the Surroundedness and Markov model.Different from the previous method of using Markov model for salient object detection,this paper combines the background prior clues and the foreground prior clues for saliency detection.Previously salient object detection algorithms often only exploit the background prior clues of the images,while ignoring the effective use of foreground prior clues.The excavation of the background prior cues in the images usually tends to simply treat all image boundaries as the background,which often results in an inaccurate saliency detection of the image which the target near the image boundary.Based on this,this paper obtains the foreground prior clues by predicting the approximate region of the salient object by using the Surroundedness.By combined with the foreground prior,this method treat the two boundaries which have longest distance away from the approximate salient region as the background prior.First,unlike the conventional algorithms using the convex hull to predict the approximate region of the salient object,this method use the Sruuondedness to predict the approximate region of the salient object for eye fixation.Then,a simple linear iterative clustering(SLIC)algorithm is used to process the original image,and the graph model of the image is established based on the superpixels.Then,the superpixels of the two boundaries which furthest from the approximate region of the salient object is taken as the virtual background absorption nodes,and the saliency value of each superpixel is calculated by the absorption Markov chain,and the initial saliency map S1 is detected.Then,the superpixels in the approximate region of the salient object is used as the virtual foreground absorption nodes,and the initial saliency map S2 is detected by the absorption Markov chain.Finally,by the fused calculation of S1 and S2,we get the final saliency map S,then using the guided filter to smooth the saliency map S.In this thesis,we test our algorithm based on two open databases of MSRA-1000 and ECSSD.The experimental results show that the proposed algorithm is better than the previous algorithms.