Seamless Texture Mosaicing For 3D Surface Models

The reconstruction of 3D models is of great importance to Virtual Reality(VR).It gets more and more widely used in medicine,movies,games,industrial modeling and visualization as well as digital preservation of cultural relics.Geometric reconstruction and texture reconstruction are the core of three-dimensional reconstruction.The purpose of geometric reconstruction is to obtain the geometric structure of the model,and the texture reconstruction provides texture information to 3D models.Texture reconstruction is mainly related to two issues,namely the registration between images and 3D models and the seamless montage of multi-view images.There are two main solutions to texturing a model with a good registration of given geometric models and multi-view images.The first solution processes the point cloud data directly by a certain color consistency standardized criteria,thus final color value of each point is expressed as the weighted average of its relating value on each visible images.Another common practice translates point cloud data into polygon mesh,and then stitches appropriate textures to each face patch,which is called labeling.The key of this problem is how to get global smooth seams.Factors causing uneven texture seams can be divided into two categories: The first includes imprecision of geometric model,registration errors,calibration errors and geometric distortion error of images center projection;the second,which is known as radiation factor,is caused by different camera angles and camera sensor changes.For the image dislocation caused by the first reason,one approach is to introduce the Markov random field model,then we can find the optimal solution by graph cuts based methods or message passing based methods.After that,we should regularize the color disparities within different texture blocks.Then some common methods like multi-band image fusion and Poisson editing algorithm can be used.In this thesis,consumer-level RGB-D camera will be used,and we will discuss the approach of seamless mosaicking of multi-textures.The main work and contributions of this thesis are listed as following:1.Proposed an approach for judging the highlight area of multi-view image andcorresponding method of highlight removal.The image of each perspective is back projected to the visible space of each triangular element.To find the highlight area within a photo,we need to calculate and compare the average pixel value of each picture area that was projected to the same triangular element.Then the highlight area is found according to the distribution of the deviation value between some single image and other images.The low frequency information of the texture near the highlight region is extracted,and Poisson fusion is carried out with the original highlight area,thereby weakening or removing the highlight effect.2.Proposed an approach for texture joint optimization considering geometric distortion and color deviation.We have made a survey on the effects of factors that may cause the texture seams inconsistency.Instead of enlarging the label set,we use a matching relation of local color bands to get shifted amount between textures.At the same time,the color of seam texture is compensated accordingly,which ensures the smooth transition between the texture seams.3.The existing algorithms of texture packing are studied.By analyzing and comparing their respective advantages and disadvantages,this thesis gives a realization considering both the storage space and rendering efficiency.Compared with the traditional texture mapping methods,the proposed scheme improves the performance of local details.