Saliency Detection Via Background Priors And Object Candidate Sets

In a complex nature scene,there are always prominent,eye-catching areas that we call salient areas.The main task of saliency detection is to accurately extract the salient target areas and output a saliency map to represent the possibility that each pixel belongs to a salient region.Because it can better help people understand images and help computers handle complex visual tasks quickly and efficiently,it has been widely used in many fields related to computer vision,such as target detection,image segmentation,video compression,and so on.In this paper,two bottom-up saliency detection frameworks are proposed based on the object candidate set and boundary background priors.At the same time,this paper proposes a novel object location prior model to optimize the saliency map.The main contribution of this paper can be summarized as followed:Firstly,based on the traditional manifold ranking model for saliency detection,this paper proposes an extended manifold ranking(EMR)model for saliency detection.EMR model integrates the background priors and regional contrasts.The traditional manifold ranking(MR)model not only makes use of the background priors,but also assumes that the non-boundary areas are all salient.This assumption is not very reasonable.Therefore,the EMR model in this paper abandons this hypothesis and redefines a new cost function.On the basis of the background prior,the salient region is found according to the prominent contrast between the salient region and the surrounding environment.Secondly,this paper extracts high-level features based on the convolutional neural network(CNN).Compared with traditional color features,the high-level features extracted by CNN has semantic information.Compared with other saliency detection methods extract high-level features based on superpixels,this paper introduces the object candidate sets from image segmentation.The shapes,textures and semantic information of object candidate sets are richer than superpixels.At the same time,for some deviations caused by object candidate regions that contain both salient areas and background areas,this paper uses the EMR method to make a correction.Thirdly,this paper proposes a novel method for calculating the position of an object.By finding the location of the saliency area,the noise area away from the saliency area can besuppressed.This article defines a cost function to find an optimal rectangular area that just blocks the salient area.In order to reduce computational complexity,this paper proposes a solution from rough positioning to precise positioning to obtain an approximate optimal solution.In the experiments,the proposed method is compared with twelve state-of-the-art algorithms on three datasets.The experimental result shows that the proposed method outperform the other twelve methods.