Research On Image Salient Object Detection Technology

With the development of technology,human beings are receiving information from all fields at any time.As an important content in image processing,salient object detection can detect and segment regions that grab interests in images,which can help image processing effectively.In this paper,we focus on how to solve the problems of localization and detailed refinement in saliency detection.The details and contributions are described as follows:Saliency detection based on the kernelized subspace ranking.We propose a subspace ranking method that takes advantage of high-level semantic information encoded with objectlevel proposals and region-based convolutional neural network(R-CNN)features.The final saliency map is generated by a weighted fusion of object proposals.The proposed method jointly learns a Rank-SVM classifier and a distance metric and can solve the problem that highdimensional samples are easy to get into the local optimum.The integrated formulation can be solved by the alternative optimization of the ranking vector and subspace ranking matrix.Compared to ranking in the primal space,subspace ranking can separate positive and negative pairs more easily and generate better performance.Salient object detection based on a stagewise refinement model.To solve the problems in deep convolutional neural networks(CNNs),that is repeated subsampling operations lead to a significant decrease in the spatial details and finer structures,here we propose to augment feedforward neural networks with a novel pyramid pooling module and a multi-stage refinement mechanism.The proposed pyramid pooling module can aggregate multi-scale context information by varied pooling kernels.Multi-stage refinement mechanism can recover the detailed information lost in the preceding stages and progressively generate full resolution saliency results.Saliency detection based on the global localization and local refinement.Effective integration of contextual information is crucial for salient object detection.The noise of Convolutional Neural Networks(CNNs)can be passed through in the most existing methods based on ‘skip'architecture.To address this problem,we proposes a global Recurrent Localization Network(RLN)which exploits contextual information by the weighted response map in order to integrate hierarchical features.Particularly,a recurrent module is employed to progressively refine the inner structure of the CNN over multiple time steps.The RLN can help accurately localize salient objects.Moreover,to effectively recover object boundaries,we propose a local Boundary Refinement Network(BRN)to adaptively learn the local contextual information for each spatial position.The learned propagation coefficients can be used to optimally capture relations between each pixel and its neighbors.The proposed three methods are qualitatively and quantitatively evaluated and compared with other state-of-the-art approaches on public salient object detection databases.Experimental results indicate that the proposed methods are effective in segmenting salient objects,achieving comparable even better performance to other techniques in terms of AUC value,PR curves,F-measure and MAE scores.