3D Model Recognition Research Based On Deep Learning

In recent years,3D model recognition has made significant progress with the help of deep learning.With the wide spread dissemination of intelligent manufacturing,3D models have not only been applied in mechatronics,architectural design,3D printing,and computer-aided design,but also been extended to various fields such as autonomous driving,AR/VR,intelligent robots,and medical image acquisition.The number of 3D models has also exploded.Effective 3D model recognition and retrieval technology can significantly improve the application rate of models in their respective fields,enhance the value of high-quality models.Therefore,3D model recognition has attracted more and more attention from enterprises and researchers.How to extract features from 3D models,retrieve the model,and manage 3D models intelligently has become hot research topics in science today.Therefore,it is of great significance to handle the massive 3D model data;how to perform 3D model recognition quickly,efficiently,and safely is a crucial problem to be solved urgently in the field of computer vision.This dissertation focuses on the deep learning based 3D model recognition algorithm,and conducts relevant researches on three core scientific issues: 1)To solve the difficulty of comprehensively feature extraction caused by the complex 3D meshes,this dissertation explores how to use the effective information of the 3D model to generate a comprehensive and accurate 3D model representation and improve the information comprehensiveness of the 3D model representation;2)To solve the difficulty ofprotecting the 3D model recognition networks caused by the vulnerability of deep neural networks,this dissertation explores how to learn the perturbation characteristics that can attack the deep neural network,generate adversarial samples that can simulate the nature perturbations,explore the reasons for being attacked from the feature distribution level,and test the adversarial robustness of the 3D model representation;3)To solve the problemthat the 3D model recognition network isvulnerable to the perturbed point cloud caused by the fuzzy boundary ofthe 3D model feature distribution,this dissertation explores how to adjust theparameters of 3D model recognition network,optimize the 3D model feature distribution,enhance the antiinterference ability of the deep neural network,and improve the adversarial robustness of the 3D model representation and network security.Based on the above theoretical system,the specific research work and contributions of this dissertation are summarized as follows:1.This dissertation proposes a 3D model recognition method based on multimodal information fusion.To solve the difficulty of the effective information extraction caused by the complex meshes of the 3D models,this dissertation designs two novel loss functions to help the network learn the correlation information from the multimodal data during training,reduce the difference in feature distribution between modalities,promote mutual learning between modalities,and reduce the time cost of the training.The correlation loss based on the correlation of different modalities.It focuses on the correlations among different descriptors generated from different structures,which aims at guaranteeing the feature consistency.The instance loss based on the characteristics of samples.It preserves the independence of each modality and utilizes feature differentiation to guide model learning during the training process.This approach can better retain the advantages of each modality,achieve the effect of complementary advantages,and finally generate a feature with more comprehensive information.2.This dissertation proposes a 3D model adversarial sample generation algorithm based on a generative adversarial network.To solve the problem of poor transferability and flexibility of adversarial samples caused by the insufficient attention to the distribution and perturbation characteristics of the 3D model recognition network,this dissertation proposes a new adversarial example generation method to explore why deep network models are vulnerable to attacks.The method first utilizes the generator to learn the representation of adversarial points effectively and adjusts the generation of adversarial points according to different point cloud inputs,which are utilized to fool the original 3D model recognition network.At the same time,the perceptual loss is proposed to improve the quality of adversarial samples by measuring the similarity between the original input and the generated samples,which can simulate the adversarial samples more realistically.The generated adversarial samples can fool the3 D recognition networks,and fool humans visually,thereby realizing the test of the adversarial robustness of the 3D model recognition network and completing the exploration of the vulnerability of the deep network model.3.This dissertation proposes an adversarial defense method to improve the adversarial robustness and security of 3D model recognition.To solve the problems of incomplete adversarial defense,high cost,and large time cost caused by the insufficient attention to the distribution and perturbation,this dissertation considers adversarial defense from the feature distribution level.The reason why adversarial examples can successfully attack the point cloud model is that the feature distribution is modified by perturbation,which misleads the output of the point cloud recognition network.This dissertation proposes to embed a perturbation generator in the 3D model recognition network,which is utilized to perturb the features during the training process,adopts the adversarial learning to optimize the parameters of the original 3D model recognition network,improve the adversarial robustness,and achieve the purpose of adversarial defense under various types of attacks.At the same time,in the process of perturbing the feature vector,this dissertation proposes to utilize the content discriminator to control the perturbation within a reasonable range to ensure that the perturbed features also belong to the original point cloud and the content is the same as the original model.Finally,this dissertation constructs a massive 3D model recognition system based on a distributed architecture with the above research results.Compared with the previous 3D model recognition systems,the system has dramatically improved in terms of data scale and retrieval efficiency,and guarantee the system stability and data security with consideration of network security.