3D Model Classification Based On EfficientNet And Multi-Feature Fusion

With the rapid development of computer-aided design,3D modelling technology and other fields,3D model has become an indispensable type of data.It is one of the important carriers after image and sound information,and has a wide range of applications in virtual reality,unmanned vehicles,industrial design and molecular biology.The number of 3D models is growing exponentially.The classification and retrieval of 3D models is the basis of research for many applications.As important part of digital geometry field,scholars at home and abroad have been focusing on this problem in recent years.This paper focuses on EfficientNet,which introduces a hybrid block attention module(CBAM)to extract view data features,and a one-dimensional convolutional neural network to extract global shape features.A 3D model classification method based on EfficientNet and multi-feature fusion is also investigated,fusing deep view features with global shape features and using them together to achieve the 3D model classification task.The main research in this paper can be divided into the following parts.1.The background of 3D model classification techniques and the current status of related research are studied.Different representations of 3D models are introduced,and3 D model classification methods based on traditional classification methods and deep learning are discussed.2.A view-based 3D model classification method is investigated to obtain the optimal angle of projection in different latitudinal directions and the best view set in different longitudinal directions,which can transform a 3D model into a set of 2D view set representations.The CBAM attention mechanism is proposed to be introduced into the MBConv module,the features of the view data are extracted using the improved EfficientNet-B0 network,the network model is optimised with data from the training view set,and the performance of the network is then tested with data from the test view set.3.The extraction of global shape features descriptor is investigated.The projected view shape is represented by D1,D2 and D3 2D shape distribution features,Zernike moments and Fourier descriptor.3D model classification method based on EfficientNet-B0 and multi-feature fusion is proposed.EfficientNet-B0 and CBAM are used to extract features from view data.1D convolutional neural network is adopted to extract global shape features.Deeper view features and global shape features are fused to exert their advantages.Finally the classification experiment is completed using fused features.4.Experimental results show that compared with traditional classification methods or other 3D model classification methods,the method based on EfficientNet and multi-feature fusion proposed in this paper can effectively improve the classification effect to a certain extent.