A Research On The Improvement Method Of IMNet For Single-view 3D Reconstruction Based On Implicit Representation

3D reconstruction is a hot research topic in the field of computational vision and computational graphics,aiming to recover 3D geometry from a given image data or video.In this paper.we study single-view 3D reconstruction IMNet(Implicit Network)network base on deep learning,mainly analyzing the effects of different attention mechanism modules,deformable convolution and adaptive graph convolution networks on the model quality of IMNet reconstruction,mainly as follows:(1)Based on the IMNet network,the attention mechanism module is embedded to improve the performance of the network and the quality of the 3D reconstructed model.Two main modules are used in this part:1.CBAM(Convolutional Block Attention Module)module:2.SimAM(Simple.Parameter-Free Attention Module)module.The results show that both modules can improve the network performance of IMNet and the quality of the reconstructed models,and the SimAM module improves the results better than the CBAM module.(2)Replace the normal convolutional network in IMNet network with Deformable Convolution(Dfcon)to investigate its accuracy variation on 3D reconstruction models.In this paper,a multi-scale progressively integrated deformable convolutional neural network is proposed for single-view 3D reconstruction.By improving the feature extraction network,the loss of semantic information is reduced and the efficiency of feature extraction is improved.The results show that by replacing the first,third and fourth layers of ordinary convolution of the feature extraction network with Dfcon.the quality of the 3D model can be improved for 61.5%of the 13 reconstructed objects.(3)To investigate the effect of Adaptive Graph Convolutional Network(AGCN)on the quality of 3D reconstruction models of IMNet networks.Inspired by the residual structure,this paper adds the AGCN module after the feature extraction layer.and inputs the obtained feature maps as nodes of the graph data into the AGCN network.which can realize the optimization of IMNet.The experimental results show that.in the process of generating a priori models,constructing a new generator by combining AGCN with Encoder and Decoder can optimize the quality of 53.8%of the 3D models of 13 reconstructed objects.The above methods are added to IMNet at the same time,and extensive experiments are conducted on 13 types of object models selected from the ShapeNet dataset.The results show that the combination of SimAM module,Dfcon module and AGCN network makes IMNet perform the best,optimizing the model quality of 92.3%of the 13 reconstructed objects and significantly improving the reconstruction of 3D models.