3D Model Recognition And Retrieval By Deep Learning

With the development of 3D sensing and digital design technology,a large number of 3D models already exist in the database and Internet of enterprises.In order to efficiently organize and reuse existing 3D models,it is necessary to automatically identify the type of the model and retrieve similar 3D models in the database.This paper draws on the breakthroughs made by deep learning methods in the field of machine vision,and studies the recognition and retrieval of 3D models by deep learning.The three-dimensional model recognition method is analyzed,and the three-dimensional model recognition method based on point cloud data is selected.At the same time,the deep neural network principle and structure for identifying point cloud data are studied.In this paper,the PointNet network model is built on the TensorFlow platform using Python language,and the model is trained using the marked point cloud dataset ModelNet40.Because,the batch normalization method adopted by PointNet is limited to the size of BatchSize,the method of group normalization is adopted in PointNet to improve the training efficiency and accuracy.Experiments show that both of them effectively improve the model's train speed and accuracy when the batch size is greater than 16,but the group normalization method is more effective when the batch size is small.In order to generate a point cloud model from the 3D model,the uniform sampling algorithm for the surface of the 3D model is studied.The algorithm first selects a triangular patch with the area size as the probability,then a three-dimensional point is sampled with equal position probabilistic in the patch,and the cycle is repeated until a three-dimensional point cloud model with a fixed number of points is sampled.Experiments show that the algorithm can effectively sample a 3D point cloud model with uniform point distribution and fixed number of points.Afterwards,a three-dimensional model recognition algorithm based on point cloud surface sampling is studied.The sampled point cloud model is input into the trained neural network after normalization,and forward prediction is used to calculate the prediction category of the three-dimensional model.This paper use C ++ and QT to realize the end-to-end 3D model recognition program,verify the feasibility of the algorithm in this paper.The retrieval algorithm based on the similarity of automatically extracted features in deep neural networks is studied.The algorithm first extracts the features of the 3D model based on the PointNet network,analyzes the feature validity using nonlinear projection,and then selects Pearson Correlation Coefficient to measure the similarity of the features after comparison.Finally,the 3D model retrieval program is designed and implemented for retrieving in the test data set of ModelNet40 3D model.The experiment proves that the retrieval results has a high similarity with the retrieved model.