Saliency Detection Via Traditional Propagation And Convolutional Network Merging

Saliency detection is extracting the most important and salient part of an image automatically by a computer.It is easy for human eyes,but difficult for computers to detect with high precision and high speed.Saliency detection can be a preprocessing step for the task of target detecion,object identification,image and video compression and so on.It can reduce the influence of redundant information,improve the efficiency of image processing and be helpful for many computer vision tasks to conduct faster and more accurately.We propose the algorithm of saliency detection via traditional propagation and convolutional network merging models,including two traditional unsupervised methods based on the disadvantages of state-of-the-art methods,which can produce two hierarchical saliency maps respectively.Then a merging model built by deep convolutional neural network is proposed to merge the saliency maps aforementioned.The main contribution of this paper can be concluded as follows:1.Saliency object is not detected completely almost by all of these methods.To address this problem,we propose Dynamic Supplementary Manifold Ranking method.It is conducted by a secondary manifold ranking,whose source image is removed partly from original saliency map to detect saliency objects sufficiently.The necessity of secondary manifold ranking is dynamically decided by the results of supplementary manifold ranking.The region belongs to background is always blurry in an image according to people's habit.As a result,blurry regions of source image are detected and corresponding locations on saliency map are depressed as background.Dynamic Supplementary Manifold Ranking and Blurry Depression are complementary to each other.2.Background prior and contrast prior are usually employed in unsupervised methods.Based on that,we propose a minimization model with constraints to detect saliency.A relationship matrix is calculated by the features of superpixels for CIE-Lab and RGB color space respectively.Superpixels with similar features are expected to own close saliency.As a result,saliency of boundaries can be propagated to the whole image according to the relationship matrix.3.The merging method of weighted average and multichannel are not suitable to the situation of this paper.We extract the first three layers of VGG-16,then add two deconvolutional layers and a convolutional layer to build our post merging network.Final saliency map can be got by the post merging network after training by the input of image,saliency maps and ground-truth map.Besides,we test our method on 10 publicly available datasets stage by stage.The results of experiments demonstrate the high performance of our traditional propagation and convolutional network merging algorithm compared with 12 state-of-the-art methods.