Research And Implementation Of 3D Object Surface Reconstruction Based On Deep Learning

In recent years,3D reconstruction technology has benefited from the tuyway brought by autonomous driving,smart city,metaverse and so on.On the other hand,thanks to the rapid development of computer vision technology,3D reconstruction has made great achievements in various fields of production and life.For example,3D CT images in medical treatment,geographic mapping,cultural relic protection,etc.In the future,3D reconstruction is essential to obtain high-quality spatial scenes in AR or VR applications.With the rapid development of deep learning,3D reconstruction based on deep learning has become a hot research direction.While current neural radiance Fields(Ne RF)based on density-based surface representation methods perform remarkably well at reproducing the appearance of objects or scenes,they cannot reconstruct high-quality surfaces.When using Marching Cubes to extract surfaces,density-based volume representations can lead to artifacts because density accumulates along the ray during optimization,rather than in isolation among individual sample points.Inspired by the advanced results in the field of deep learning,this paper improves the implicit surface representation method based on Truncated Signed distance Function(TSDF)and the color rendering method based on brightness to solve the artifact problem,and constructs a hybrid scene representation method by merging the two representation methods,so as to design and propose a 3D reconstruction network model based on deep learning.The proposed model consisted of two Multilayer perceptrons representing shape and brightness respectively.The density-based scene representation method in Neural Radiance Field(Ne RF)was replaced by a hybrid scene representation method,and the model was extended to use depth maps from RGB-D sensors.In addition,a camera pose optimization technique and a forward scene expansion method are proposed.The camera pose optimization technique can compensate for the misalignment in the input data and improve the overall reconstruction quality.The forward scene extension method can extend the applicability of the scene representation network to the real scene.The 3D reconstruction network proposed in this paper is fully tested on both real and synthetic data,and shows excellent reconstruction effects.Compared with the advanced 3D reconstruction models in recent years,the 3D reconstruction network proposed in this paper has the highest Intersection over Union(Intersection over Union)score,normal consistency and F-score score in seven scene datasets.