Image 3D reconstruction is widely used in reverse engineering, artificial intelligence and other fields. It is a hot research topic based on deep learning to reconstruct a 3D model using a single image. The article first reviews the research status of 3D reconstruction of single image, focusing on 3D-R2N2 based on voxel expression, PSGN based on point cloud expression, Pixel2Mesh based on monolithic grid expression and AtlasNet based on multi-piece triangle mesh expression. The algorithm analyzes the problem of different expressions of different tasks and output models through experimental comparative research.