Deep learning-based multi-view 3D reconstruction is a method for reconstructing a 3D scene from image data from multiple viewpoints using deep learning techniques. Existing deep learning-based 3D reconstruction networks suffer from reconstruction residuals and missing point clouds at the edges, resulting in poor accuracy and completeness metrics of the reconstruction. The reasons for this are the lack or loss of local texture feature information of the image during 3D reconstruction and poor feature matching. Our method firstly combines the ability of the feature pyramid to detect features and targets in the image at different scales, and calculates the self-attention to the image features at each stage of the feature pyramid, and then sums up the results of the calculation with the output of the FPN. Secondly in order to improve the quality of edge reconstruction, this paper performs edge feature extraction on the depth map and designs a loss function for the depth features to improve the completeness of edge reconstruction. Finally, qualitative and quantitative experiments on the DTU dataset prove the effectiveness of our method, and qualitative experiments on the Tanks and Temples dataset demonstrate the generalization of our method.