Research On Visual Saliency Detection And Salient Object Segmentation

With the rapid development of the society,the breakthrough of science and technology and the increasing popularity of the Internet,the means of image acquisition are more and more convenient and flexible,thus the amount of image data acquired is also increasing explosively.Compared with the massive and growing image data,computing resources are limited.How to use the limited computing resources to automatically and efficiently analyze and understand image content is a huge challenge in the field of computer vision.Visual attention mechanism enables people to quickly find and process some significant or interesting area of a complex visual scene,whose core task is visual saliency detection and segmentation.Based on the existing research,this dissertation conducts in-depth research on visual saliency detection and salient object segmentation.The main contents and contributions can be summarized as follows:1.The saliency maps obtained by the current frequency-domain-based visual aliency detection methods generally do not have clear boundary.We propose a visual saliency detection algorithm fusing global and local information based on nonsubsampled contourlet transform(NSCT),which can well detect and segment the salient object.The NSCT has properties of multiresolution,localization,directionality,and anisotropy,which makes it able to describe the image details effectively and approach the smooth contour precisely.In this dissertation,a bottom-up salient object detection approach fusing global and local information based on NSCT is presented.Images are first decomposed by NSCT and the high-frequency coefficients are categorized and optimized accordingly to get better representation.Then feature maps are obtained by performing the inverse NSCT on these optimized coefficients.The global and local saliency maps are generated from these feature maps.In the end,the final saliency map is computed by fusing the global and local saliency maps together.Experimental results on MSRA 10K demonstrate the effectiveness of the proposed method.2.Most saliency detection methods rarely consider the image illumination influence on the detection results,and some previous saliency detection methods based on frequency domain generally use the high-frequency of images for analysis without considering the low-frequency information.In order to solve these problems,a coarse-to-fine saliency detection framework based on NSCT is designed.Firstly,the influence of scene illumination on saliency detection is discussed,and Retinex theory is used to improve the traditional saliency detection methods.On this basis,the Retinex-corrected saliency detection method on the low-frequency components is used to measure the coarse saliency.The high-frequency feature maps are enhanced by the coarse saliency,and the fine salieny map is calculated from global and local perspectives.This method can utilize both low-frequency components and high-frequency components effectively and suppress the influence of illumination.Qualitative and quantitative evaluations on ASD,DUT-OMRON and MSRA-10K datasets verify the feasibility and effectiveness of the proposed method.3.For the salient object segmentation in natural images with intensity inhomogeneity and complex background,two level set segmentation methods combined with saliency information are proposed.One is a regional level set segmentation method combining intensity correction and saliency information.Retinex-correction is to suppress the influence of intensity inhomogeneity caused by non-uniform illumination and enhance the following saliency detection;The corrected image and saliency information are embedded in the regional level set energy function,and the segmentation is completed by curve evolution.The other is a level set segmentation method combining the saliency-embedded region information and edge information.First,the Retinex theory is introduced to correct the saliency information extraction,thus suppress the effects of intensity inhomogeneity on saliency detection and subsequent segmentation.Second,the Retinex-corrected saliency information is embedded into the level set energy function due to its advantage which makes a foreground object stand out relative to the backgrounds.Combined with the edge information,the boundary of segmentation result will be more preciser and smoother.Experimental results show that the proposed method is robust and effective.4.The basis of the wavelet transform is lack of anisotropy,and it is not as good as the nonsubsampled contourlet in sparse approximation of object contour.However,the level set method can achieve the purpose of approaching the obejct contour through curve evolution,and has good expansibility.Therefore,the level set method is used to compensate for the limitation of wavelet transform approximating contours and a salient object segmentation based on wavelet transform saliency information and level set method is designed.First,the GBVS algorithm is used to calculate the image saliency,and the initial contour is automatically located by using CV level set method on the GBVS saliency map.Then the image is decomposed by wavelet transform,and the high-frequency components are reconstructed to obtain the feature maps,which are then enhanced by GBVS saliency.The saliency is measured from global contrast and local contrast.Combined with the initial contour and the saliency information based on wavelet transform,the curve evolution is carried out according to the level set method and the salient object is segmented.The experimental results on the database can verify the effectiveness of the proposed method.