Research On 3D Model Retrieval Method Based On Convolutional Neural Network

Thanks to the rapid development of the Internet and computer hardware and software technology,and the growing maturity of 3D modeling technology,compared to text and 2D images,3D models not only contain the information of the object itself,but also the model.Spatial information,the relative distance between adjacent objects,can more vividly display a three-dimensional world that fits the human eye more closely.In this context,the classification and retrieval of 3D models has important commercial application prospects.The early 3D model retrieval method was based on text.Because this method is too personally subjective,it is difficult to apply to complex and variable 3D models,and then gradually turned into content-based retrieval.This type of approach is usually implemented by two methods: model-based retrieval and view-based retrieval.Modelbased retrieval extracts features directly from the 3D model data and retrieves them in the database.However,the prior art mainly focuses on the overall feature.The function of local retrieval is not complete,and the ability to retrieve objects from heterogeneous objects is insufficient.The most important thing is that the features obtained directly from the three-dimensional model have strict requirements on the equipment.In response to the above issues,we chose a view-based retrieval method.This paper focuses on the three-dimensional model retrieval method based on convolutional neural network.Firstly,we propose a classification model based on multi-view convolutional neural network.In the classification model,we use a multi-view twodimensional image to represent a three-dimensional model,each of which independently passes through the first convolutional neural network,and all branches share the same convolutional neural network parameters.Each view produces a feature that uses the View-pooling pooling layer to aggregate the features of the multi-view 2D map to generate a simple and efficient 3D shape descriptor as the real input to the second convolutional neural network.The unnecessary noise information is discharged by the PCA dimension reduction,so that the features better represent the threedimensional model and improve the retrieval efficiency.Experiments show that the proposed method can achieve better classification results,and achieve 91.00% and 81.5%,respectively,in classification accuracy(accuracy)and retrieval accuracy(mAP).In addition to this method,we also propose a method for three-dimensional shape retrieval based on multi-layer convolutional neural network,which has more comprehensive information,in which two low-resolution paths are added to the network,which can capture the visual appearance and reduce Invariance between images.Experiments have shown that the classification and retrieval performance is further improved,and the classification accuracy(accuracy)and retrieval accuracy(mAP)are increased by 0.7% and 0.6%,respectively.