3D Model Retrieval Method Based On Deep Network

In recent years,with the rapid development of virtual reality technology and Internet technology,the number of 3D models has increased dramatically.At the same time,3D models gradually become a new type of multimedia data.All of these advantages make 3D model technology being used in more fields than ever,such as Virtual Reality(VR),medical research,Computer Aided Design(CAD),3D film animation,industrial product design,mechanical manufacturing and many others.The amount of 3D models in the Internet is huge and the growth rate is growing.How to quickly and accurately retrieve the required target model from the Internet,while saving time and resource consumption has become an urgent need in many fields.Therefore,3D model retrieval technology came into being.Nowadays,content-based 3D model retrieval technology has received extensive attention.Although this method has a shorter development time and there is still room for improvement of it,it produced satisfied performance.This method implements the retrieval by extracting features from the model and then performing similarity measurement.Compared with the text-based retrieval,content-based retrieval can exclude subjective influences and has strong objectivity and superiority.This paper focuses on the view-based 3D model retrieval algorithm,and proposes two algorithms: 3D model retrieval based on CNN-LSTM network and 3D model retrieval method based on convolution recurrent neural network.Firstly,the 3D models are preprocessed,then the 2D view sequences are extracted,then the sequences are sent to the convolutional network for bottom layer feature extraction.Then the LSTM neural network or recurrent neural network take these extracted features as input for optimization learning.After the fully connected layer,the similarity measure is performed.The final output is the probability representation of each 3D model.The methods presented in this paper were evaluated on four popular datasets: the NTU database,the PSB database,the ETH-80 database,and the MV-RED database.And the experimental results are compared with the classical algorithm,which shows the effectiveness of our methods.