Research On Classification Method Of Cultural Relic Fragments Based On Hierarchical Structure And Point Cloud Augment Network

With the development of virtual repair technology and 3D scanning equipment,the digital protection of cultural heritage is advancing by leaps and bounds.Many cultural relics can be digitized using existing scanning equipment and stored in computer equipment.However,since most of the unearthed cultural relics have been damaged during excavation,further processing and restoration are required.The location identification of cultural relic fragments is a key step in the restoration of cultural relics.The classification method can identify and classify the cultural relic fragments according to the preset position category.The efficient classification method can lay a solid foundation for the subsequent splicing restoration work and accelerate the restoration and reorganization process of the cultural relics.Aiming at the technical bottleneck of the current digital protection of cultural relics,which is unable to extract effective feature information and low classification accuracy.In this paper,the hierarchical structure design with attention mechanism and the multiresolution and multi-feature fusion strategy are applied to the classification of cultural relic fragments.The main research contents are as follows:(1)A three-dimensional cultural relic fragment classification method based on hierarchical structure design is proposed.First,the key points of the point cloud are obtained through down-sampling operation,and the local area is constructed with each key point as the center.The feature information of each local area is extracted through the multi-layer perceptron,and the attention module is constructed to learn the influence factor of each feature channel in the local area.According to the degree of importance,different feature channels are assigned corresponding weights to enhance the features that have a higher effect on the classification task,and inhibit the features that have a lower effect.Then,the center loss is introduced on this basis.The joint cross entropy loss and center loss are used to update and optimize the point cloud network model.Experimental results prove that the method can fully extract the local information of each key point,and maintain a high classification accuracy without increasing a lot of calculation.(2)This paper proposes an improved augment cultural relic fragment classification framework.Mainly to improve the augmentor in the original augment framework,downsampling the original input of the augmentor multiple times to obtain multi-resolution point cloud input.The multi-layer perceptron is used to extract point cloud features of different resolutions,and the shallow features and deep features of the point cloud are combined to generate the global features of the point cloud.Fusion of multi-resolution and multi feature point cloud features to improve the learning ability of the augmentor.The augment framework is guided by classification results,and uses adversarial learning strategies to jointly optimize and update the parameters in the augmentor and the classifier.Among them,the augmented samples generated by the augmentor learning are more adaptive than traditional methods.Experimental results prove that for different point cloud data sets,the improved augment framework has a good improvement in the classification performance of different classification networks,and is higher than the classification accuracy of the original augment framework.