| According to the theoretical basis of computer vision such as polar geometric constraint and plane scanning algorithm,the multi-view stereo matching 3D reconstruction restores the spatial structure of the object by imaging different camera perspectives in a certain scene.In the field of three-dimensional reconstruction,this method has take advantages of low cost and wide application range.In recent years,it has been widely used in the fields of auto driving,medical image processing,preservation of virtual reality,cultural relics protection and so on.The deep learning method trains the input sample through the neural network,which has the characteristics of learning the deep semantic information and global features of the input sample.In recent years,many researchers have transferred deep learning methods to the field of three-dimensional reconstruction.Based on deep learning technology,some scholars have built an end-to-end 3D reconstruction network model(Multiple View Stereo Network(MVS-Net),which has achieved better 3D reconstruction results,and the accuracy and integrity of the reconstruction point cloud have been improved.However,there are two obvious problems with rebuilding point clouds: first,the object boundary and background part in the point cloud are not effectively distinguished,and there are many edge noise and background noise;Second,the reconstruction point cloud has poor accuracy,low completeness and average visualization effect in the texture-rich part.In order to solve the above problems,this paper studies the multi-view 3D reconstruction network model MVS-Net based on deep learning,improves the steps of feature extraction and cost regularization.In addition,Cascade Cost Volume is used in the steps of differential homography transformation,the large depth intervals and few depth intervals are used to predict the depth range on the down-sampled low-resolution images,and then the predicted results are used on the high-resolution images that are not down-sampled.More accurate depth prediction is obtained.Mainly work as follows:(1)In the stage of feature extraction,this paper combines U-Net and improved DRE-Net to extract richer semantic information in the picture.The U-Net structure completes semantic segmentation,which solves the problem that the object boundary and back shadow in the reconstruction point cloud are not effectively distinguished.The DRE-Net structure with improved attention mechanism obtains rich semantic information in images,which solves the problem of low accuracy and poor integrity of reconstructing point clouds in texture-rich areas.The next sampling on the low resolution image with large depth interval and less depth interval prediction depth range,and then forecast results to high resolution image by sampling,obtain more accurate depth prediction.(2)In the stage of cost regularization,this paper combines the spatial attention module and the 3D CNN structure to filter edge noise and isolated noise more effectively.In the symmetric codec structure of 3D CNN,this paper uses the spatial attention module to complete the pixel-level comparison of the cost bodies of the same scale,and applies the spatial information weight grid to the decoding part to filter the cost bodies multiple times to solve the problem of more edge noise and background noise in the reconstruction point cloud.The experimental results show that the proposed model can filter the edge noise and backgroud noise more significantly,and the reconstruction in the area of sparse texture is more complete,and the reconstruction more in the area of rich texture is explicitly,and the Overall reconstruction visualization effect is better.Compared with the representative models in the field of 3D reconstruction in the last three years,the model in this paper achieves excellent reconstruction results on the open source indoor data set DTU.Good reconstruction results have been achieved on open source outdoor data set Tanks and Temples. |
