Research On 3D Model Recognition Algorithm Based On Deep Learning

The number of three-dimensional models is increasing day by day.However,manually labeling three-dimensional models requires additional human resources and is less efficient.3D model recognition accuracy based on artificial design features is low.Although convolutional networks have good performance in various visual applications such as image classification and target detection,but the introduction of a large number of pooling layers will cause the loss of feature space information.CapsNet saves the spatial attributes of features through vector neurons,which is more robust to model changes.Therefore,Capsnet can be used in 3D model recognition to learn the feature information of 3D models.The main research contents of this article are as follows:(1)In the task of 3D model recognition,for the problem of the loss of feature space information due to the excessive number of pooling layers,according to the characteristics of the capsule network,a 3DSPNCapsNet(3D Small Pooling No dense Capsule Networks)for identifying 3D models is proposed.First,the three-dimensional model is voxelized into data types that can be input into the network;secondly,the new network structure is used to extract more representative features;finally,DRL(dynamic routing algorithm with length information)based on dynamic routing algorithm(DR)is proposed to optimize the iterative calculation process of capsule weights.Experiments on Model Net10 show that,compared with 3DCapsNet and Voxnet,this method achieves better recognition results,the average recognition accuracy rate in the original test set reaches 95%.(2)In the rotation recognition task of the 3D model,In order to verify the ability of the capsule network to recognize the spatial position and angle transformation of features,the3 D model is first rotated and then voxelized to expand the training set of different rotation angles to add the rotation information learned by the network.And also rotate the test set.Second,the voxelized rotation model of each angle is visualized to prove the effectiveness of the rotation.Finally,several algorithms are used on the rotation training sets with different rotation angles and tested on a rotation test set for all rotation angles.The experimental results show that the recognition rate of 3DSPNCapsNet for the nonrotated model reaches 96%,and the average recognition rate of the rotated models at other angles reaches 81%,which has good recognition ability for the three-dimensional model and its rotation.