Research On 3D Model Semantic Extraction Technology Based On Deep Learning

With the development of 3D modeling technology,the amount of 3D model data has increased sharply,and it is extremely difficult to manually process and analyze a large amount of 3D model data.In this context,it is of great significance to automatically analyze and process 3D models with computer.Based on the artificial intelligence method of deep learning,research the semantic extraction technology of 3D models,let the computer autonomously understand the overall semantics and part of the semantics of the 3D models,corresponding to the shape recognition technology and semantic segmentation technology of the 3D models respectively.This thesis research the semantic extraction technology of 3D models,including feature extraction methods,shape recognition technology and semantic segmentation technology of 3D models.The specific work is as follows:Aiming at the feature extraction methods of 3D models,five current representative 3D model feature extraction network structures are analyzed:PointNet and PointNet++based on point cloud,VoxNet and O-CNN based on voxel,and MeshNet based on mesh.Shape recognition and semantic segmentation experiments were carried out on the public data sets ModelNet40 and ShapeNet,respectively,and the feature extraction capabilities of five network structures to 3D models were analyzed.For the shape recognition of the three-dimensional model,shape recognition algorithm(Mesh-Skeleton-Net)combining skeleton features is proposed.Taking the skeleton of the 3D model as the data,skeleton feature extraction neural network is proposed.Combining MeshNet's spatial feature extraction network,structural feature extraction network and mesh convolution module,shape recognition neural network fused with skeleton features is proposed to achieve the shape recognition of the 3D model.The experimental results on the ModelNet40 dataset show the effectiveness of the skeleton feature extraction network,and Mesh-Skeleton-Net has achieved better performance in shape recognition.For the semantic segmentation of 3D models,context-reinforced octree convolutional neural network algorithm(CR-O-CNN)is proposed.The octree-based convolutional neural network is introduced into the context-reinforced network,and the learning process of the context features in 3D model is modeled as Markov decision process,and the process is optimized by Asynchronous Advantage Actor-Critic algorithm to learn deep context features,and gradually improve the segmentation results of the 3D model.The experimental results on the ShapeNet dataset show that CR-O-CNN has achieved better semantic segmentation performance.