Research On High Precision Multi-view Stereo Reconstruction Algorithm Combined With Shading Cues

Image-based surface reconstruction has a wide range of applications in industries such as medical,education,game/movie,live broadcast and interactive events etc.After several decades of development,significant improvements have been achieved.However,there are still problems unsolved,such as reconstruction for fine structures,surfaces with specular reflection,untextured areas with complex lighting environment,and real-time interactive scene reconstruction etc.As an important part of the machine to understand the surrounding world,such research is necessary and urgent.Recent algorithms can either smooth out some details while improving the completeness,or it is difficult to simultaneously improve the reconstruction completeness of the untextured areas while restoring details.To solve problems mentioned above,we propose a joint optimization framework combining superpixel-based multi-view stereo depth completion with shading constraint,which can restore the surface details and improve the reconstruction completeness.Since the existing MVS(Multi-View Stereo)algorithms are difficult to restore a reliable surface shape in the untextured region,a superpixel-based depth completion and spatial propagation is proposed to further improve the reconstruction completeness of the untextured regions in this thesis.Meanwhile,to tackle the problem that high-frequency details on the surface are easily lost for objects with spatially varying albedo,a global continuous overall illumination lighting model is proposed.And the geometric shape obtained at each step is used to jointly optimize ambient lighting,object surface reflection and geometric vertices.While improving the completeness of the reconstruction,we can recover more high-frequency details on the surface.To verify the effectiveness of the proposed algorithm,we use Middlebury and ETH3 D dataset to do evaluation experiments.As seen from the experimental results,the proposed algorithm greatly improves the reconstruction result both qualitatively and quantitatively compared with existing open source algorithms.It shows that the algorithm in this thesis can enhance the recovery ability of high-frequency details and improve the completeness robustly.