3D Model Classification Based On Deep Learning Of Multi View Feature Fusion

3D models are widely used in all walks of life.Compared with 2D data,3D data has stronger expression ability and richer information.The emergence of a large number of 3D models has brought about difficulties in data annotation and classification of 3D models.Traditional 3D model classification methods need manual features extracted manually to classify 3D models.There are many problems in this method.Firstly,with thousands of data,it takes a lot of manpower and resources to annotate and extract manual features from 3D models.Secondly,manual features have a lot of subjective judgments,and can not obtain higher-order features with more information,resulting in low accuracy of 3D model classification.With the rapid development of deep learning,view-based 3D model classification has become a research hotspot.Existing methods have achieved good experimental results,but there are still the following problems.Firstly,current method is used for network training,the feature extraction of the difference and importance information between different views and within the view is insufficient,and it is difficult to obtain representative 3D model feature descriptors.Secondly,current 3D model classification methods are validated on the datasets whose initial poses are aligned.However,in practical applications,the poses of 3D models are unknown,resulting in obvious performance degradation of a non-aligned 3D models.Finally,the massive 3D models not only diverse within class,but also similar between classes,which seriously affects the classification accuracy of 3D models.In order to solve the above problems,three 3D model classification methods are proposed in this paper.The main work is:(1)A 3D model classification method with multi-view feature fusion is proposed to solve the problem of difficulty in obtaining representative 3D model feature descriptors.During data processing,the method uses a tight field of view with a black background combined with Phone to render 2D views.During network training,the hybrid domain attention mechanism is added to the view feature extraction network to obtain the main features inside the view.Then fuses the view features to obtain the global features of the 3D model and inputs the global features into the view weight learning network with channel domain attention,giving different weights to the views according to their importance to the 3D model,forming a representative 3D model feature descriptor for 3D model classification and improving the classification accuracy of the 3D model.(2)A pose non-aligned 3D model classification method is proposed to solve the problem of low accuracy of pose non-aligned 3D model classification.This method employs graph convolutional neural network(GCN)to learn the spatial relations between views,and uses the preset camera positions as the vertexes in the graph structure.Moreover,the timing feature extraction network and the attention network are used to further improve the effect of GCN.Experiments show that the method has a much higher classification accuracy than existing methods in the case of aligned 3D model poses.It also has higher classification accuracy in the case of non-aligned 3D model poses.(3)A 3D model classification method based on comparative learning is proposed to solve the problem of the existence of a large number of similar models in a large amount of 3D model data.In this method,training is divided into sample discrimination stage and classification stage.In the sample discrimination stage,the same category 3D models are set as positive samples and other category 3D models are set as negative samples,and the sample features were constrained using contrast loss to map the samples onto the same single-centre unit hypersphere in space to obtain a good semantic representation space for 3D model classification.In classification stage,by fine-tuning the network parameters,the network model is migrated to the classification task to complete the 3D model classification,which further improves the accuracy of 3D model classification.