Research And Implementation Of 3D Reconstruction Based On Tensor And Learning

Due to the broad application prospects of multi view 3D object reconstruction tasks in AR,surveying,and medical fields,multi view 3D object reconstruction has attracted widespread attention from many researchers.In recent years,research on multi view 3D objects has found that the task of multi view Iterative reconstruction mainly focuses on reducing reconstruction cost and improving reconstruction quality.This article proposes two methods around reducing reconstruction costs and improving reconstruction quality,as follows:The first method in this article aims to propose a deep tensor learning method for accelerating reconstruction and reducing model size while ensuring network performance.Specifically,this paper chooses to CP decompose the full convolution network.Because the parametric model of the full convolution network is suitable for tensor decomposition,it can effectively reduce model parameters,improve the processing speed of convolution,learn low rank information,and use expansion convolution to replace the original 3*3 convolution,increase the Receptive field,capture multi-scale information,and learn sparse information.The second method of this paper aims to propose a multi-scale Receptive field network structure construction feature extraction module to improve the reconstruction performance.At the same time,combined with the depth separation convolution and expansion convolution,the multiscale Receptive field network is constructed.Starting from feature extraction,the learning ability of the feature network is.improved.The main work of the paper is:1.This paper proposes a method based on tensor decomposition and sparse learning to reconstruct the convolution network in the 3D object reconstruction task.It uses the low rank constraint of tensor decomposition to achieve the purpose of parameter dimensionality reduction,and uses the void convolution to replace the original convolution,so that the Receptive field during convolution learning is larger,and sparse learning is carried out through the void convolution,which together with the low rank structure of tensor decomposition forms the low rank plus sparse learning.Qualitative and quantitative experiments were conducted,and the results showed that compared to the benchmark method,when selecting a rank of 3,the decomposed 2D CNN parameters were compressed to 1%of the original parameter quantity.The accuracy was improved compared to the original baseline model,and the integrity performance indicators were also reconstructed well.The qualitative results showed that the reconstructed 3D model had less noise,and this method ensured the reconstruction quality while reducing the model size.2.This paper proposes to build a multi-scale Receptive field network based on cavity convolution,design the expansion rate to build a zigzag convolution network based on HDC principle,form a multi-scale Receptive field,then introduce nonlinear structure through deep separation convolution,deepen the network without introducing additional parameters,improve the learning ability of the feature network,and conduct quantitative and qualitative experiments.The results show that the accuracy and integrity of this method and the overall score index are better than the benchmark method,Effectively improving the reconstruction results.3.Designed and implemented an online 3D object reconstruction system,integrating the method proposed in this article into the system,enabling users to freely log in and upload images for 3D object reconstruction.