Research On 3D Model Recognition And Retrieval Based On 3D Convolutional Neural Network

With the continuous progress of scientific and technological information technology,the number of 3D models has increased exponentially,playing an important role in industrial production and daily life.How to classify and manage the models has become a matter of widespread concern and research in academia and industry.The continuous improvement of computer hardware performance has made the convolutional neural network,which has been slowly progressing due to computational performance bottlenecks,once again achieve high-speed development,especially in the field of 2D images recognition.The particularity of the 3D model structure makes its recognition and classification process different from traditional images.Therefore,this paper considers applying 3D convolutional neural network to the classification and retrieval of the 3D model.The main tasks are as follows:A weight optimization integrated convolutional neural network model is proposed and applied to the classification and recognition of 3D model.This method first obtains the depth projection view of three angles of front view,top view and side view of the 3D model,so as to retain the spatial information of the 3D model to the greatest extent.Then uses the improved VGGNet to train the depth projection images of each view and output its prediction probability value for the category of the 3D model.Finally,a weight calculation algorithm is used to assign weights to the prediction probability value of the depth projection view and weighted integration,thus completing the final classification of the 3D model.A convolutional neural network model based on adaptive weighted integration is proposed.Based on the entropy,the model assigns appropriate weights to the depth projection view features of the front view,side view and top view of the 3D model extracted from the deep learning model.Then,the features and corresponding weights are weighted and integrated into a new fusion feature,and the weights are used to increase the proportion of the features with good initial classification performance in the fusion features.Logistic regression is used to classify the fusion features to complete the classification and recognition of the 3D model.An integrated 3D convolutional neural network model based on weight optimization is proposed.This method first renders some 2D images of a 3D model at a specific position from the top view,and outputs them in the form of video in order.Then tries to describe the voxelized 3D model with the data structure of multidimensional binary matrix.Finally uses the improved 3D convolutional neural network model to train these two forms of data separately and outputs the probability of each category to which the 3D model belongs.According to the algorithm,the two results are given appropriate weights,and then the weighted integration of the predicted probability value of the preliminary classification is carried out to identify the classification of the 3D model.Experiments show that this method can effectively improve the classification accuracy.