Research On RGB-D Image Salient Object Go-Detection

The human visual attention mechanism can help all the people quickly find out the most-interesting part in the complicated environment,or we can say the most-salient part.Based on above all,the salient object detection makes use of computer to simulate human visual attention mechanism,so as to have a fleet search on the salient object region in the images.Now salient detection as a pre-process for images has had a widespread use on various fields,like segmentation of images or videos,location and weak-supervised learning etc.At present,salient detection is mainly divided into two directions,salient detection for single image and co-salient detection for multiple images.The different between normal salient object detection and co-salient detection lies in that the previous one aims at detecting the salient region of a single image,while the latter one is aimed at detecting the common salient region of a group of images which can be divided into the common salient region and the non-common salient region.According to different detection images,salient detection for single image can be divided into salient detection for single RGD image and salient detection for single RGB-D image.As well,co-salient detection can be divided into co-salient detection for multiple RGB images and co-salient detection for multiple RGB-D images.This thesis proposes a co-salient model,which can be applied to RGB-D image.Firstly,this thesis uses the single existing salient image as the initial image,extracts superpixels-level color and depth feature of the image,and since deep learning network has a salient function on computer visual field,this thesis makes advantage of a deep learning network to extract out superpixels-level high-dimensional semantic feature of the image,then combines with the initial saliency,color,depth and high-dimensional semantic feature,so as to find out superpixel's consistency among the images;secondly,this thesis introduces and makes some improvement of cellular automaton model that can takes the superpixels with strong consistency as the new neighbor among the images,and adds the comparison of depth value when calculating the weight value of superpixel;finally,this thesis utilizes the co-cellular automaton model to optimize the single existing salient image so as to get the final result.In the above work of this paper,four types of features,namely color,depth,initial saliency and high-dimensional semantic features,are extracted manual clues for co-detection.However,some groups of images is sensitive to depth information,some groups of images is sensitive to color information on account of the great difference among images in different groups,one unified cue may not apply to all image groups.Therefore this thesis expects to have a co-salient detection under the help of individual sensitive characteristic selected by the feature selection,so as to reach a better effect.This thesis unites particle swarm optimization(PSO)of intelligent optimal computation field,proposes a co-salient detecting algorithm for RGB-D images based on PSO feature selection cluster.Firstly,according to the initial image of RGB-D single salient images,select candidate salient region,then extract 24-dimensional superpixel-level features from these candidate salient regions;then utilize the clustering algoritlhm based on PSO feature selection to generate a base salient image;finally,this thesis introducing EMR algorithm to optimize the base salient image,and gain a final result.This thesis makes a verification and assessment of the two RGB-D co-salient object detection methods proposed from two public benchmark RGBD Cosegl 83 and RGBD Cosa1150 with other state-of-the-art.Shown from the experiment result,this thesis has a vide superiority in the assessment of PR curve and F-measure value.