Retrieval Modes Of 3D Models Based On Generative Adversarial Network

With the continuous improvement of computer graphics processing ability and 3D modeling technology,more and more multimedia data appear and are widely used by users.As a new generation of multimedia data after voice,text and video,3D model data have been widely studied and applied in the fields of games,virtual reality environment,medical diagnosis,and computeraided design,which promotes the emergence and rapid development of 3D model retrieval technology.However,there are still some challenges in 3D model retrieval technology,which hinders the development of 3D model retrieval technology.Content-based 3D model retrieval mainly achieves model to model retrieval by extracting the characteristics of 3D model itself.However,the complex spatial structure and shape information of 3D model make it difficult to fully represent the feature information in many practical applications.In addition,because text cannot accurately describe a 3D model,the core problem(semantic gap)of cross-media retrieval of 3D model and text has not been well solved.In view of the problems in content-based 3D model retrieval and cross-media retrieval of 3D model and text,this paper proposes the corresponding methods respectively.The main work and innovations of this paper are as follows:(1)This paper proposes a group-pair deep feature learning method for 3D model retrieval.The method mainly includes three core stages: view feature extraction,group-pair feature learning and feature fusion.In this paper,an improved Convolutional Neural Network(CNN)is used to extract the view descriptors,and the maximum pooling and average pooling of the pooling module are used to aggregate the view descriptors to reduce the loss of effective information in 3D model.Because different types of features contain different 3D model information,this paper employs supervised autoencoder and multi-label discriminator to further mine the latent features and category features of 3D model.In this paper,two features are fused in series to form a more discriminative shape descriptor,and a Margin Center Loss is defined to further improve the retrieval performance of 3D model.The experimental results in Model Net10 and Model Net40 datasets show that the proposed method is superior to other methods.(2)This paper proposes a method of text retrieval model based on multi-modal auxiliary classifier generative adversarial network.In this method,the sharing network projects 3D model and text data into a common subspace to obtain the same semantic feature representations.The discriminator can distinguish the mode and category of feature representation to improve the feature learning ability of sharing network.Meanwhile,this paper defines a neural network optimization method based on structure-preserving loss to improve the accuracy of model retrieval.In addition,an autoencoder is introduced to construct multi-modal auxiliary classifier generative adversarial network with autoencoder based on the former network.It mainly exploits the data reconstruction characteristic of autoencoder to further reduce the semantic gap between 3D model and text and enhance the cross-media retrieval performance of 3D model and text.The experimental results in XMedia Net dataset show that two networks both have good retrieval effect.This paper takes feature learning as the research goal.The feature representation learning capability of CNN,autoencoder and GAN is employed to extract deep features of 3D model.Through the intelligent fusion of different types of features,the efficient 3D model single-modal retrieval and cross-modal retrieval of 3D model and text methods are given respectively to realize retrieval modes of 3D models based on GAN.