Research On Multi-view Stereo In 3D Reconstruction Based On Multi-scale Cost Volumes

In the deep learning based multi-view stereo method in 3D reconstruction,the results of feature extraction and the effect of generating cost volumes directly affect the accuracy of the reconstruction.In order to address the problems of general feature extraction effect and insufficient correlation between multi-scale cost volumes in current methods,this paper proposes a multi-view stereo network model based on multi-scale cost volumes.First,in order to obtain more complete and accurate feature information of the input image,a double U-Net feature extraction network model is proposed,while the output features are structured in a cascade from coarse to fine according to three different scales.Secondly,in order to make the extracted features pay more attention to the detailed information of the desired target,a network using attention mechanism for feature enhancement is proposed to post-process the extracted features.Finally,in the cost volume regularization stage,this paper proposes a preprocessing network model for sharing information of multi-scale cost volumes.This model separates the information within small-scale cost volumes and passing it to the larger-scale cost volume for fusion,and performing depth map estimation from coarse to fine.So that the reconstruction accuracy and completeness are substantially improved.In this paper,we conducted experiments on the DTU dataset,and compared with the original method,the three main metrics of accuracy(Acc.),completeness(Comp.)and overall(Overall)were improved by about 18.3%,8.4% and 13.5%,respectively,which were greater than other deep learning-based methods,and in several other minor indicators have also improved in varying degrees.For each improved part,experiments are also conducted separately to verify the effectiveness of the improvements.The experimental results show that the method proposed in this paper effectively improves the effect of feature extraction and feature enhancement,increases the interconnection between multi-scale cost volumes,and has a certain improvement in reconstruction accuracy compared with the original model and other methods,which also verifies the real effectiveness of the proposed method.There are 30 figures,9 tables and 64 references in this paper.