Research On Multi-View Stereo Matching Algorithm Based On Deep Learning

Multi-view 3D reconstruction technology recovers the 3D spatial information of the object from the 2D image of different perspectives and generates the 3D point cloud of the target object,which has the advantages of low cost and strong applicability,and has a wide range of applications in virtual reality,security monitoring,urban planning,automatic driving and other fields.The process of multi-view 3D reconstruction includes input of multi-view images,sparse reconstruction to calculate camera parameters,and multi-view stereo matching to generate dense point clouds.At present,the existing multiview stereo matching method can achieve good reconstruction effect in an ideal Lamberian scene,but for weak texture areas,due to the difficulty of local feature extraction,it is difficult to calculate the accurate pixel depth value in the stereo matching process,resulting in poor 3D reconstruction effect and prone to point cloud void problems.This thesis focuses on the above problems,and the main research content is as follows:(1)Aiming at the difficulty of feature extraction in weak texture areas,a multi-scale feature extraction D-RFP network is proposed.Based on the recursive pyramid structure,the network repeatedly extracts and fuses the feature information in a recursive way,which can extract accurate and rich multi-scale feature information.In the D-RFP network,a deep separable convolution operation is introduced,which greatly reduces the amount of computation and parameters of the network and ensures the operation efficiency of the whole model.Comparative experiments show that the completeness of point cloud reconstruction of DTU dataset is improved by 0.12 mm.(2)Aiming at the problem of point cloud voids,this thesis studies PatchmatchNet and proposes an improved PatchMatch module for accurate depth estimation.In the deep propagation diffusion process,a more robust base diffusion mesh is used to obtain better neighborhood depth assumptions.In the process of matching cost body construction,the convolutional attention mechanism is introduced to calculate the similarity grouping weights so that the computing targets focus on important feature channels and feature regions.Finally,the adaptive spatial filtering operation is used to optimize the depth map.Comparative experiments show that the completeness of point cloud reconstruction in DTU dataset is improved by 0.08 mm.(3)Based on the proposed multi-view 3D reconstruction algorithm and CesiumJS3 D earth rendering engine,this thesis designs a 3D dense reconstruction and visualization system,which can realize the reconstruction of 3D dense point cloud and efficient rendering and visualization on 3D earth.