Optimization Technology For Multi-view Stereo Algorithm

With the development of the new subject of computer vision,3D reconstruction technology has became a popular research direction.It has been widely used in medical diagnosis,film industry,the 3D games and many other fields.The main purpose of 3D reconstruction is to recover the target scene.Multi-view stereo is one of the main methods of obtaining 3D models.Among these methods,a 3D reconstruction method which is proposed by Furukawa(PMVS method)is better in both of the accuracy and the efficiency.This thesis provides a deep research on the PMVS and proposes an improved PMVS algorithm.The main works of this thesis are following:(1)The reconstruction results obtained from PMVS algorithm are easily to produce holes due to the uneven distribution feature points.This thesis presents an adaptive 3D reconstruction algorithm based on an improved feature point detection method.New method firstly separate the Harris operator's response matrix into blocks.Then we compare the maximum response value in each block with the mean response value of the response matrix.Finally,we select the feature points in different situations.In this way,we solve the problem of how to set the appropriate threshold.The cluster phenomenon of feature points is relieved.And there are fewer holes in the results of the 3D reconstruction.The time of the reconstruction is also saved.The experiment results indicate that our method is effective.(2)The reconstruction results obtained from PMVS algorithm are easily to run a long time due to the expand step.This thesis presents a new method which firstly make the sparse patches into dense patches,and then do the expand step.By using the NCC algorithm and the use of the gradient of disparity method to deal with outliers.In the premise of ensuring the quality of matching points,using the quasi-dense matching method to expand the sparse patches.Spreading the sparse patches to dense patches.Then do the follow-up operations.The experimental results show that,after improvement,the time-consuming of the expand step is saved.The speed of the 3D reconstruction is improved.(3)The reconstruction results obtained from PMVS algorithm are easily to run a long time due to the traditional BFGS method.This thesis presents a new 3D reconstruction algorithm based on an improved BFGS method.This paper changes the modified matrix of the quasi-Newton method.This change guarantees the global convergence conditions of the BFGS method and a same result of the 3D reconstruction,with less computing time.Through the change of the quasi-Newton method,which indirectly reduce the time-consuming of the 3D reconstruction.The experiment results indicate that our method is effective.