Multi-view 3D Reconstruction Based On MVSNet

Multi-view 3D reconstruction technology is a mainstream research topic in the field of computer vision,and the technology has important research value in the fields of Autopilot and medical diagnosis.Traditional methods use manual feature extraction,which makes the reconstruction process more complicated and the accuracy of the reconstructed model is not high.Compared with traditional methods,the process of deep learning-based methods is simpler and greatly improves the model accuracy and reconstruction efficiency.Therefore,this paper provides an in-depth study of MVSNet,a multi-view 3D reconstruction network based on deep learning,which is mainly as follows.1.In response to the problems of low local integrity and noise interference in MVSNet,this paper proposes an adaptive channel attention 3D reconstruction network SE-MVSNet.Due to the phenomenon of partial incompleteness and excessive noise in MVSNet,its reconstruction effect is not ideal.Therefore,this article adds channel attention(SE)between every two cost and probability volumes of the same scale in the network,adaptively adjusting the weight of noise related channels through attention mechanism,thereby reducing noise interference,improving the accuracy of depth estimation,and obtaining better 3D reconstruction results.The reconstruction performance was evaluated on the DTU and Blended MVS datasets,and our method achieved0.379 mm and 0.414 mm in completeness(Comp)and overall quality(Overall),respectively.The results show that the proposed method not only performs better than advanced multi view 3D reconstruction methods such as MVSNet and P-MVSNet,but also significantly improves the visualization results.2.The adaptive channel attention 3D reconstruction network SE-MVSNet has the following problems:(1)In the process of feature extraction,the global Semantic information and local texture information of the image are not effectively used;(2)The channel attention SE module only focuses on channel dimension attention,ignoring spatial dimension position information,which is not conducive to the network’s more comprehensive learning of cost volume features,resulting in low accuracy and incomplete information in SE-MVSNet.To alleviate the above issues,this paper proposes a three-dimensional reconstruction network BFC-MVSNet that integrates multi-scale features and mixed attention.This method is based on MVSNet network,using Bidirectional Feature Pyramid Network(Bidirectional Feature Pyramid Network,B-FPN)for feature extraction,while utilizing low-level and high-level features of the image to obtain multi-scale fused features in a more effective manner;In the regularization of the cost volume,we add mixed attention(Convolutional Block Attention Moudle,CBAM)to improve the accuracy of depth estimation by strengthening the expression of important features of the cost volume.The results on the DTU dataset indicate that the reconstruction performance of our method is superior to advanced multi view 3D reconstruction methods such as MVSNet and P-MVSNet.Meanwhile,experimental results on the Tanks and Temples dataset show that the proposed method has great generalization performance.