Abstract: In order to improve the accuracy of 3 D reconstruction from single image,a deep learning based neural network is proposed by improving the Pix2 Vox network for 3 D reconstruction from single image.Firstly,multi-scale connection and channel attention mechanism are added to the Pix2 Vox network structure to retain multi-scale information and enhance key feature learning.Secondly,a threshold calculation module is proposed to implement the threshold setting method adapted to different categories and optimize the threshold value.Finally,a fusion loss function is proposed to fuse the structural loss and the class loss of the model to reduce the influence of unbalanced data and class differences on the reconstruction effect.The experimental results show that the average IoU of the proposed network is 0.670 in the 13 model categories of ShapeNet dataset,indicating that better 3 D reconstruction performance can be achieved than using the Pix2 Vox and other networks.