Classification And Retrieval Of 3F Models Based On Multi-Feature Fusion

In recent years,the rapid development of computer technology makes it more convenient to obtain data information in 3D models and store model information.Nowadays,3D models have been widely used in the fields of architecture,medical treatment and industrial manufacturing.At the same time,the number of models is growing rapidly.How to retrieve the required models from the numerous 3D models and carry out design and processing is crucial to increase the utilization rate and improve the work efficiency.Therefore,3D model recognition and retrieval technology has become a research hotspot.With the wide application of deep learning technology in two-dimensional images,some excellent results have been achieved,and the model retrieval technology of deep learning comes into being.However,most of the existing methods focus on the learning of a single feature,and it is difficult to carry out a comprehensive learning of the model features.Therefore,this paper conducts relevant research on 3D model classification and retrieval based on multi-feature fusion: to improve the multi-view network,improve the network classification and retrieval performance and obtain the model view features;Using attention mechanism to improve the classification effect of three-dimensional point cloud neural network;The feature fusion network MFFAUTL based on autoencoder was constructed,which fused the features of view,point cloud and voxel to improve the classification and retrieval performance of 3D models.The model retrieval system is built.The main contents are as follows:(1)A 3D model retrieval method MVCNNG based on group similarity learning and view grouping is proposed to solve the problem of inaccurate model classification and poor retrieval results in MVCNN due to the loss of adjacent view similarity information and poor view feature fusion effect during maximum pooling.The first is to use the group similarity learning of views to improve the MVCNN framework in the maximum pooling when losing the similarity information between adjacent views.Secondly,a view grouping model is constructed based on metric learning,which is used to group views and integrate view features.Experimental results show that the classification and retrieval effect of this method is better than that of the original network.(2)The three-dimensional point cloud classification algorithm based on attention mechanism is discussed,and the channel attention model TSE is used to optimize Point Net,which improves the classification performance.In addition,by converting the grid model into a voxel model,the voxel characteristics of the three-dimensional model are obtained.(3)A network framework MFFAUTL based on auto-encoder multi-feature fusion is proposed.According to the characteristics of the encoder,a multi-feature fusion network is constructed to integrate the features of view,point cloud and voxel.The new feature combines the information of the view and the spatial information of the three-dimensional model.Experiments show that this framework is more effective than single feature in 3D model classification and retrieval.(4)A 3D model retrieval system is developed based on python language.Through this system,model retrieval,display,format conversion and model management can be carried out.Finally,the system is retrieved and tested,and the functional effect is good.In conclusion,this paper focuses on the topic of 3D model classification and retrieval based on multi-feature fusion,proposes the improvement and fusion scheme of relevant feature learning,improves the classification and retrieval effect of multi-feature fusion network,and builds a 3D model retrieval system.