Research Of Salient Object Detection Based On Prior Knowledge Guidance

In recent years,a large number of scientific literature have explored the operating mechanism of machine vision systems from different perspectives in order to make machine vision capable of human visual attention mechanisms.Among them,saliency object detection is an important application of computer vision.The purpose of this application is to quickly and accurately find the most prominent area of human attention at first glance in the image,namely the salient object area.In addition,the detection of saliency targets is also applied in image segmentation,image classification,image recognition,image retrieval,target relocation,visual tracking and other fields.Saliency detection approaches can be categorized as two main kinds:bottom-up and top-down.Bottom-up usually exploits low-level cues such as features,colors and spatial distances to construct saliency maps.In contrast,top-down is task-driven and requires supervised training with manually labeled truth map.We use the bottom-up on salient object detection.A large number of extension algorithms have been proposed based on the bottom-up and top-down ideas.Some use sliding windows or elastic edge frames to obtain the saliency map,which can lead to the approximate position of the salient object,and can't effectively highlight the outline of the object.In view of the above problems,we propose a saliency detection method for center rectangle composition priori.Firstly,from the perspective of foreground,salient objects are distributed in the center of the image and its surrounding areas.A central rectangle is used to cover the Region,consider the super-pixels on the rectangle as salient seeds and obtain the saliency map based on the center composition line.Then,under the same theoretical guidance,we calculate the salient values of the intersection points based on the center rectangle composition.Then,based on the spatial compactness,the salient values based on the spatial compactness relationship distribution are obtained.Finally,we use the method of Gaussian fusion to put the three together to get the final saliency map.It is just based on the idea of the priori salient detection of the center rectangle composition,a multi-scale self-search saliency detection method combined with the rectangular diffusion is proposed.The method is based on the perspective of foreground,using a plurality of rectangular boxes to cover the original image,and using the super-pixels on the rectangular box as the salient seeds for the manifold ranking to obtain the salient values based on the rectangular diffusion.Then,from the perspective of background,all the super-pixels are weighted based on the color features,and the seed nodes that may be the foreground are excluded from the background.Then,the background seeds are used for sorting in manifold ranking,and the salient values are obtained by self-searching.Finally,the salient values of the rectangle diffusion and self-searching are linearly fused to obtain the final salient object.With the advent of stereo camera such as Microsoft Kinect,saliency object detection for RGB-D image is attracting more and more interest.Each RGB-D image can be decomposed into a color image with R,G and B color channels and an aligned gray-level depth map.Many researches focus on how to use depth information to supplement color cue for saliency detection.Some people use a simple linear fusion of the salient results obtained from the RGB map and the depth map,which does not emphasize the role of deep information and does not work well.Some people still treat the RGB-D RGB of the image using commonly used RGB algorithms,obviously neglecting the fact that the human visual system operates in a real 3D environment.We propose saliency object detection for RGB-D image from 3D perspective.It regards object as three dimensional structures,and redefines boundary conception in RGB-D image,and regards space boundary including top,down,left,right,front,back plane in real 3D environment as background.It incorporates 3D compactness feature,in which salient objects typically have 3D compact spatial distributions,into color and depth feature to express similarity among super-voxels and applies manifold ranking by six boundary planes to generate six saliency maps,and then integrates them to get the RGB-D saliency map from background view.In the end it refines saliency map by high confident salient seeds from foreground view.This thesis for RGB image processing,is MSRA-1000,CSSD,ECSSD,THUS-10000,SED a total of five data sets were tested.Test results show that the current state-of-the-art algorithms have greater advantages.This thesis is for RGB-D image processing,is NLPR RGBD1000 dataset tested.Test results show that the current state-of-the-art algorithms have greater advantages.