Saliency Detection Based On Manifold Regularized SVM Model

Image saliency detection aims to remove redundant information and obtain the regions of interest in the image.As the preprocessing step of image,it is widely applied in computer vision area with multiple applications,such as object detection.Here,we propose a semi-supervised saliency detection algorithm-saliency detection by manifold regularized SVM model.All superpixels in the image are fully utilized to train the saliency detection model,thereby,reducing the training costs and improving the efficiency and precision of saliency detection.Besides,we simultaneously take the characteristics of local learning and global learning into account,the local manifold regularized methods often ignore the completeness of salient objects especially when there is the low contrast between the foreground and background,while the global regularization can discover the long-range semantic structure,but ignore some detail information of the object.therefore,we simultaneously construct the local and the global manifold regularized SVM models,integrate saliency results,making saliency detection results more perfect.First,we over-segment an input image into a set of superpixels and produce a series of object proposals.We learn a dense affinity matrix,based on which object proposals are exploited to compute an initial saliency map,thereby determining pseudo labeled foreground and background superpixels.Second,respectively construct a local regularizer and a global regularizer,and train two manifold regularized SVM models using pseudo labels to predict saliency values for all superpixels.Third,the local and global regularized results are integrated in a principled way.Finally,update the labeled samples to re-train the SVM models for saliency refinement,thus getting the final saliency results.Extensive experiments on five benchmark datasets indicate that our algorithm performs better than the other saliency detection algorithms.Furthermore,we show that the manifold regularized SVM model can be easily applied to some existing saliency models for significant performance improvement.