Research And Implementation Of 3D Reconstruction Algorithm Based On Deep Learning

With the development of reverse engineering technology,3D reconstruction has been integrated into many industries such as medical imaging,cultural relic digitization,surveying and mapping.Using a standing 3D scanner or a hand-held 3D scanner for 3D reconstruction is not only bulky and expensive,but also has high requirements for application scenes,which seriously limits the application range of 3D reconstruction technology.Image-based 3D reconstruction technology is to obtain the depth image of the scene by imitating the vision of animals,and then carry out 3D model fitting to achieve the effect of reconstruction.Compared with the traditional reconstruction method,this technique has the advantages of light equipment,low cost,low application requirements,and better reconstruction effect.With the development of deep learning,3D reconstruction based on deep learning has become an important focus in the research field of 3D reconstruction based on images.In this paper,by analyzing the existing 3D reconstruction methods based on deep learning,the existing network model has serious memory consumption,low efficiency,and can not meet the needs of large scene reconstruction.An high efficiency multi-view Stereo network(H-MVSNet)model is proposed.H-MVSNet builds lightweight feature extraction modules,regularizes with improved gated recurrent unit(GRU)modules,and refine depth maps with Levenberg-Marquardt(L-M)layers.The network model is optimized to reduce memory consumption and improve network efficiency.Meanwhile,H-MVSNET adopts the method of fusion of original image sequence and predicted rough depth map to improve the generation quality of depth map and improve the effect of 3D reconstruction model.In the test of DTU data set,the accuracy error between the 3D model generated by H-MVSNet and the original model was 0.327 mm.It only took 0.44 s to calculate a depth map,and the memory consumption was as low as 2.46 GB.The Temples and Tanks data set tests were slightly better than the previous network.In this paper,the proposed method is quantitatively and qualitatively evaluated.Experimental results show that H-MVSNet can effectively improve the accuracy and accuracy of 3D reconstruction,improve the computational efficiency,meet the needs of large-scale 3D reconstruction,and effectively improve the quality of lightweight neural network in 3D reconstruction applications.The ablation experiment further shows that the model is more effective in solving the problems of severe memory consumption and low network utilization.