3D Model Retrieval Based On Convolutional Neural Network

At present,view-based 3D model retrieval is a research hotspot,widely used in machine vision,3D model retrieval and other fields.This method first projects3 D models into 2d view sets,and then uses deep learning techniques to classif y and retrieve them.However,there are still some problems with current methods.First of all,current model classifiers do not make full use of model pose information.Secondly,most of the current methods use more views when performing 3D model retrieval,which reduces retrieval efficiency.Finally,the feature dimension extracted by deep neural network is high and the retrieval efficiency is low.In order to solve the above problems,this paper proposes targeted solutions,mainly including the following aspects:Firstly,the classification of 3d models based on single view and pose information is proposed,to solve the problem that most models do not consider pose information,it also greatly reduces the number of views required for categorization.The method first uses all projected views to train a baseline system for feature extraction,then,using the classification accuracy obtained from the baseline system,a linear regression model was trained to select a single view.In this method,a baseline system is obtained by using all projected views to extract features,then a linear regression model is trained by using the classification accuracy obtained from the baseline system to select a single view,and then different model classifiers are trained by using views of different poses.Finally,a 3D model classification algorithm is proposed based on our method.In the 3D model classification,a classifier is selected according to the algorithm for model classification.Experiments show that the method pro posed in this paper can effectively use the pose information of 3D models to improve the accuracy of3 D model classification.Secondly,a 3D model retrieval method based on representative view is proposed to solve the problem of too many views and too low retrieval efficiency.The method first trains a picture classification network based on Res Net50 network framework to extract picture features.Then,the method of K-means is used to cluster the obtained projected view to obtain the representative view.Finally,a 3d model retrieval strategy is proposed,which can terminate the retrieval process in advance and effectively improve the retrieval efficiency of3 D models.Experiments show that the representative view selection method based on K-means can reduce the number of matching views,and the retrieval strategy in this paper can also significantly reduce the number of matching,so as to improve the retrieval efficiency.Finally,a depth hashing based feature extraction and retrieval space partitioning for 3D models is proposed to solve the problem of high feature dimension extraction by rolling neural network.This method first designs a deep hash network and trains the hash layer function.Then the fuzzy matrix of the classification is used to set up the retrieval space and the index is established in the retrieval space according to the hash characteristics.Finally,the index is used to retrieve the 3d model,and it can greatly improve the retrieval efficiency and accuracy.Experiments show that the feature extraction based on depth hashing compresses the feature dimension,and the indexing based on hashing feature greatly improves the efficiency of model matching.The above contents are optimized from different stages of 3D model retrieval.First,the projection view is grouped to make use of pose information effectively.Then select the representative view of the projection view,which can effectively reduce the redundant view.Finally,the deep hash model is used to compress the features extracted by deep learning and build an index.Experiments show that the method in this paper can improve the accuracy and efficiency of 3D model retrieval.