Visual Saliency And Nonparametric Statistics Based Scene Segmentation And Reconstruction

In recent years,with the rapid development of the Internet,image and video data have increased dramatically.Scene segmentation and reconstruction based on these data has always been one of the research hotspots and difficulties in the field of the computer vision.This dissertation focuses on the research of scene segmentation and scene reconstruction based on visual saliency and nonparametric bayesian statistical methods.It follows the technical route from scene static segmentation to scene dynamic reconstruction,and the main contributions of this dissertation are listed as follows:1)For a single colorful image with depth information,a scene semantic object segmentation method based on multi-modal consistent saliency is proposed.This method combines multimodal information and a variety of prior information to effectively solve the problem of extracting the significance of similar color and texture effects between salient objects and background,and to extract more accurate significance,which is an important and effective feature for scene segmentation.This method includes a simple and effective adaptive multimodal over-segmentation method,which can accurately segment the whole scene.Based on these accurate over-segmented regions,this method calculates the saliency combined with the global contrast and the prior information of the foreground and background.By extensive verification and comparative experiments,the results show that this mothed can achieve better performance,and is superior to other methods in accuracy and recall rate.2)On the basis of obtaining the salient semantic segmentation results of foreground objects in a single image,this dissertation proposes a method of scene sequential image reconstruction based on nonparametric probabilistic semantic object association,which is used to associate the semantic objects among the sequential images.In the process of scene dynamic reconstruction,this method can be used to match and correspond to the same objects at different times to solve the problem of object level data association during the scene dynamic reconstruction process.Firstly,we introduce the basic principle and graph model of nonparametric Bayesian Dirichlet process model,then connect the graph model with the elements in the process of scene dynamic reconstruction,and this dissertation studies the relationship between graph model parameters and scene dynamic reconstruction,so as to effectively realize the association of semantic objects.In the experimental part,the simulation and comparative experiments shows that the proposed method is advanced.We verify the effectiveness of the proposed method in real scene dynamic reconstruction.3)In addition to the sequential images,for the nonsequential images,this dissertation also proposes a nonsequential image reconstruction method based on nonparametric probability key-frame extraction,which improves the efficiency of the reconstruction in the process of struction-from-motion.During the scene dynamic reconstruction process,this method can filter out fewer keyframes,which can represent scene information better,to alleviate the problem of low computational efficiency in large-scale scene reconstruction.After keyframe selection using distance dependent Chinese restaurant process model,redundant images with little visual angle changes and little contribution to 3D reconstruction are automatically removed,which can greatly reduce the number of images in the process of 3D reconstruction and improve the efficiency of reconstruction.The experimental results show that the proposed method can effectively reduce the time consumption of 3D scene reconstruction of multi view disordered images.By the comparison of simulations and experiments,the feasibility and effectiveness of three proposed methods are verified.Finally,it summarizes the whole content of the dissertation,and looks forward to the problems and directions that need to be explored in the future.