In the field of deep learning 3D reconstruction, CVP-MVSNet and CasMVSNet have the memory consumption problem compared with MVSNet, which significantly reduces the memory consumption of the model when processing high-resolution images and the accuracy error of reconstructed point clouds, but the integrity error of the point clouds is large. In order to solve the above problems, a multi-view 3D reconstruction network based on dilated attention and depth optimal correction is proposed in this chapter, which is named DA-MVSNet. DA-MVSNet takes CasMVSNet as the benchmark network, and introduces a feature enhancement network composed of parallel dilated convolution and attention module that integrates the idea of deep separable convolution, aiming to enhance the global feature capture ability of the reconstruction network to the input view and improve the integrity of the reconstructed point cloud. At the same time, in order to prevent the feature enhancement network from extracting too much view non-relevant background information and causing the degradation of the accuracy of the reconstructed point cloud, an optimization correction of the depth map based on the idea of nonlinear least squares is introduced in the output part of the network, aiming to improve the accuracy of the output depth map. The experimental results show that the DA-MVSNet network model has the best quality and good comprehensive performance when running on the indoor scene dataset DTU, and it also runs on the large dataset Tanks Temples in the real outdoor scene, showing good generalization ability.