Research On Improvement Of Multi-view Stereo Matching Network Based On Deep Learning

The purpose of Multi-View Stereo(MVS)3D reconstruction is to reconstruct the dense geometry of a 3D object from a series of images,corresponding camera positions and internal parameters.Traditional geometric shape restoration and reconstruction methods mostly use hand-designed feature descriptors to determine the correspondence between pixels in different images,and apply engineering regularization to restore 3D point clouds,which makes the reconstruction effect in smooth areas such as weak texture and specular reflection poor.With the maturity of deep learning,MVS algorithm based on neural network enhances the ability of capturing object features and shows obvious advantages in benchmarking.Based on the multi-view 3D reconstruction method of deep learning,a 3D Cost Volume was constructed,which was regularized by the 3D convolutional neural network.Finally,the depth map was returned from the probability body,and the depth information from the multi-view was fused into a 3D point cloud.Although this method achieves better performance than traditional methods in benchmarking,it does not scale well to high resolution scenarios,and the running speed and memory allocation are also difficult to meet most application scenarios.In order to enhance the practicability of the algorithm,this paper improves the regularization and depth range estimation methods based on MVS.A high-precision multi-view 3D reconstruction algorithm GRU-MVSNet based on deep learning is proposed.The gated recurrent unit(GRU)module is used to replace 3DCNN for cost body regularization.A new GRU Range Estimation Module(GREM)is proposed,which can dynamically determine the depth ranges of each pixel in the depth map and obtain more accurate depth values,further improving the quality of the three-dimensional reconstruction.In the final test of DTU data set,the overall error is0.312(mm),and the output time of a depth map 0.79(s).Performs better than most networks on the Tank&Temple dataset.The effectiveness of the proposed module is further proved by the ablation experiments.In general,GRU-MVSNet effectively improves the accuracy and integrity of 3D reconstruction.