Research And Implementation Of 3D Model Recognition Method Based On Semi-supervised Deep Learning

Three-dimensional(3D)model recognition is one of the research hotspots in computer graphics and computer vision.At present,deep learning has become the mainstream of 3D data recognition,but most of the existing methods are based on supervised learning,which requires a large number of accurately labeled 3D data.Moreover,labeling 3D data is time consuming,and a large number of easily obtained unlabeled data are not fully utilized by existing algorithms,which leads to the difficulty in improving the effect of 3D model recognition algorithm and weak adaptability.To solve this problem,this thesis studies how to use semi-supervised learning method to build 3D recognition model by using a large number of unlabeled data to improve the effect of 3D model recognition.Firstly,this thesis proposes a semi-supervised method of collaborative training(co-training)to train the 3D recognition model,which selects two basic classifiers based on different views for co-training.The unlabeled shape is pseudo-labeled to participate in the later iterative training process.Through the experiments,the proper classifiers needed for the construction of this model are selected,and the applicability and effectiveness of the co-training architecture for 3D shape recognition tasks are proved.Secondly,this thesis proposes an uncertainty-aware consistency loss to make the two basic classifiers learn the consistency information of unlabeled data in different views.To reduce the negative impact of single consistency loss on model learning,this thesis introduces the uncertainty of the model on the basis of consistency loss to dynamically adjust the consistency loss function,thus reduces the negative impact on the whole model because of the performance gap between basic classifiers when learning different unlabeled data.Through relevant experiments,the positive effect of uncertainty-aware consistency loss on model learning of unlabeled data is verified,which can significantly improve the learning effect.Thirdly,this thesis proposes two methods of data pseudo-labeling to improve the quality of the pseudo-labeling.One is soft-labeling method and the other one is uncertainty-based threshold controlling method.Through the experimental comparison,the final choice is to use uncertainty-based threshold controlling method to improve the quality of pseudo-labeling,and further improve the final learning effect of co-training model.Finally,a 3D model recognition system is designed and developed.The system provides services based on the browser client,and adds an intermediate layer in the server to provide model call interfaces for the client and shield the specific model operation details.At the same time,the system provides four functions: shape recognition,file analysis,data labeling and model training.The system provides common operations used for the process of 3D model construction and training,which is convenient for users to participate in the process of 3D shape recognition model construction and training.