Research On Deep-learned Scalable Multi-View 3D Reconstruction Method

This paper proposes a scalable multi view 3D reconstruction method based on depth learning,which can process image datasets with different scales and resolutions.With the whole process of multi view reconstruction,the scale problem has different meanings for reconstruction by taking different images of the same target from different distances.Large-scale images can overcome weak texture,provide rough estimation of location,and small-scale images have richer texture information and improve reconstruction accuracy.In the past,multi view 3D reconstruction methods mostly used image data from the same scale,which limited the data source.In order to overcome the difficulties of reconstruction based on multi-scale image,the algorithm proposed in this paper has the following innovations: 1.In different stages of multi view 3D reconstruction,different scale representation methods are introduced,and the scale adaptability of the original algorithm is improved;2.In multi view depth estimation network,multi-scale cyclic neural network,semi global regularization,and geometric perception are proposed 3D CNN and other methods to improve the multi-scale adaptability of existing networks.In this paper,multi view 3D reconstruction is divided into four processes: depth map reconstruction,point cloud reconstruction,mesh reconstruction and texture reconstruction.Firstly,the definition of local voxel scale is given,and the image set used to create depth map is selected by using image filtering rules considering scale information.The algorithm is also used to select multiple depth maps for comparison to remove outliers.Next,the depth map reconstruction algorithm is improved by combining the depth learning theory,so that it has the ability of multi-scale matching and regularization:the improved feature extraction module Mini UNET is more concise and can maintain the accuracy than the commonly used Uni Net;the multi-scale cyclic neural network and semi global valence regularization are proposed,so as to expand the receptive field of the cyclic neural network to the global;the cascaded variation is adopted Depth optimization,updating the depth value iteratively,so as to expand the solution space.Then,in the stage of point cloud reconstruction and grid reconstruction,the method in this paper carefully considers the combination of local voxel scale and traditional multi view 3D reconstruction,so the generated model has the best available resolution: the local part of point cloud can contain small details of millimeter level,and the whole city area can be included as a whole,which depends on the scale of the original image;the local scale is considered for grid Degree of the three-dimensional diloney map cut generation,after the generation of denoising.Finally,texture reconstruction uses Markov random field to create reasonable texture.In this stage,gradient intensity integral of texture block is used to describe scale.In this paper,on the basis of scale adaptability,the parallelism of the algorithm is improved and the space complexity of the algorithm is controlled as much as possible.This paper trains DTU data set and compares with popular algorithms.Compared with colmap,the reconstruction speed of depth map is at least 20 times faster than that of colmap,which is equivalent to that of p-mvsnet.In order to verify the generalization ability of the algorithm,this paper trains on DTU,tests on eth3 d and tanks and temples,the accuracy is 7 % 11 % higher than p-mvsnet;in addition,this paper also trains on blendedmvs,and tests on private datasets,and obtains better visual effect.