Visual Saliency Objects Detection With Graph Based Semi-supervised Learning And Ranking Models

With the development of network technology and the explosive growth of data volume,how to extract useful information from these massive data and carry out effective application has brought great challenges to researchers.Visual attention is an important mechanism for Human Vision System to identify saliency parts of a scene.Humans can easily process such visual information and quickly "interesting" or"important" areas by filterling unuseful information easily.Researchers pay more and more attention on the study of vision that how to use computer to get the "interesting"or "important" areas by simulating and learning from the biological cognitive mechanism of human vision system.The saliency objects detection is to locate"interesting" or "important" areas by using computer from simulation of human visual system.Visual saliency objects detection results can be used as a pre-processing process in many other research fields and have been widely used in many research fields such as image retrieval,image segmentation and target recognition.Compared with traditional machine learning technology,semi-supervised learning can make full use of a small amount of labeled data and its own characteristics of a large amount of unlabeled data to help improve the predictive performance of learning.The graph based semi-supervised learning method model has been widely used in saliency objects detection and has obtained good detection results.Such as random walk model,manifold ranking model and so on.Although the existing methods have made some progress,there still have many problems to deal with.This thesis mainly analyzes the existing visual saliency objects detection methods based on graph semi-supervised learning and proposes several models,and verifies them on multiple datasets.The main research contents of this thesis include:saliency objects detection based on global and local consistency model,saliency objects detection model based on prior regularization ranking,saliency objects detection by multi-scale cooperative ranking methods and saliency objects detection method based on multi-view learning.The main works and innovations of this thesis include:First of all,a global and local consistency computation model is proposed to calculate the saliency objects.By the study of traditional saliency objects detection method based on the random walk theory,we propose a global and local consistency model.The model based on the Markov chain of random walk model and calculated by the average absorption time from transient node to absorbed node,the global information can be captured from the image,then the local information can be captured by using manifold features,the two kinds of information application in the are used in global and local consistency model.By using boundary information and foreground information as query nodes,the saliency value can be obtained by two-stage process.Experimental results show that this method can effectively obtain the structure information of the image,obtain more accurate saliency map,and improve the accuracy of saliency objects.Secondly,we propose the prior regularization saliency detection method based on single-layer graph and multi-layer graph.The task of image saliency objects detection is to obtain the saliency value of each pixel of the image.The traditional manifold ranking method can well capture the manifold structure of images and obtain better detection results,to get full use of the prior information,we propose to add a prior information as a regularization term to improve the model and propose a saliency objects detection method based on single-layer graph.Then,we do more research on the construction of multi-layer graph and proposed a novel method for saliency detection.Experimental results on multiple datasets show that our two proposed methods have better detection results.Thirdly,we propose a multi-scale cooperative novel method for saliency detection.To fully exert the influence of different scales of saliency detection,first of all,we use the pyramid model and get the images of different scales,and then by super-pixel image segmentation for different scales super-pixels,and then use our proposed multi-scale cooperative ranking method on different scales of super-pixels at the same time,and use different scales to the consistency of the ordering of values for the final result,through the experiment on the multiple datasets,proved that the proposed method has better detection results.Fourthly,the saliency objects model of multi-view learning based on graph is proposed.Compared with single-view point data,multi-view data can obtain the essential information of data more comprehensively and systematically.Multi-view learning is widely proposed for in-depth understanding and analysis of multi-view data.In the computation of visual saliency objects,the image can be described by different features such as color,texture and shape to obtain multi-view feature data.In order to capture different visual feature information to obtain more accurate saliency objects,we propose a saliency optimization method based on multi-view learning.Firstly,the existing saliency methods based on graph are summarized and classified,and a general saliency optimization framework based on graph is proposed.Then,the single graph model is extended to multi-view case,and the multi-view saliency optimization model based on graph is proposed.Finally,an implementation of the general model is given,and an effective updating algorithm for solving the model is derived.Experimental results on multiple benchmark data sets show the effectiveness of the model.