Research On Multi-view-based Geometric Deep Learning Model Of 3D Data And Its Adversarial Attack

With the rapid development of multimedia technology and 3D reconstruction technology,people begin to pay attention to more complex 3D data.3D data contains more abundant object information,and with the improvement of computer hardware equipment and computing and storage capacity,the amount of data is gradually increasing,and the processing of 3D data is becoming more and more important.Therefore,the deep learning of3 D data has been paid attention to this field is called geometric deep learning,which has become a research hotspot in recent years.Among them,more and more researchers pay attention to classification and retrieval.In this thesis,through the investigation of 3D shape recognition algorithm,summarizes the existing research methods of multi view,proposes a new method based on multi view depth panorama embedding feature learning to classify and retrieve 3D objects;in order to verify the anti-interference ability of geometric deep learning model,this thesis proposes three kinds of view based learning models(DeepPano?MVCNN and 6V-DeepPano)attack.(1)Aiming at the problem of classification and retrieval of three-dimensional objects,this thesis designs six different angles to project three-dimensional objects,which are the main view,the top view,the left view,the right view,the bottom view and the back view angle,these six views can basically accurately classify the three-dimensional model.In order to reflect the geometric position information of three-dimensional data in views,this paper uses cylindrical projection method to obtain depth panoramas containing 2.5-dimensional information from these six views,and proposes a convolution neural network based on six view depth panoramas.Experiments show that the proposed method is effective,and the classification and retrieval results on modelnet40 dataset are better than those using 12 perspectives in multi view methods.(2)In order to verify the anti attack ability of geometry deep learning model,this thesis designs a general anti disturbance experiment on 3D shape retrieval system,and designs three kinds of learning models based on DeepPano?MVCNN and 6V-DeepPano.The experimental results show that the attack effect of the proposed method against these models is significant,and the existing geometric deep learning model has the vulnerability of poor anti-interference ability.Among them,the anti-jamming ability of the learning model with geometric information is better than that of the two-dimensional image,which indicates the importance of the geometric information in the deep learning model,which is more robust and safe for geometric deep learning model design,and provides a defense scheme to improve the security of the retrieval model.