Research On Virtual Restoration Method Of Cultural Relics Based On Point Cloud Processing

As the bearer of history,cultural relics reflect the social activities and economic foundation of a certain historical period.They play a crucial role in clarifying the development of civilization and understanding history.Due to natural and man-made reasons,many cultural relics are being destroyed or even disappearing,and the protection of cultural relics is urgent.The traditional method of cultural relics restoration is not only slow and costly but also easy to cause secondary damage.Therefore,the use of emerging technologies to repair damaged cultural relics is imminent,especially the use of deep learning algorithms to complete the 3D model of cultural relics.To improve the quality of point cloud completion,reduce the influence of noise and outliers,and eliminate redundant and highly correlated features,simplification of point cloud data is also an urgent problem to be solved.In order to improve the quality of point cloud simplification and keep enough feature information while ensuring integrity,this paper proposes two point cloud simplification algorithms based on feature extraction.To apply the deep learning algorithms to the completion of the cultural relic point cloud,this paper constructs a cultural relic point cloud dataset,and a multi-scale point cloud completion network is proposed.The main work of this paper is as follows:(1)A scattered point cloud simplification algorithm(SPSA)is proposed.Firstly,the algorithm finds and preserves the boundary information of the point cloud model.Then use the FPFH(fast point feature histograms)feature descriptor to calculate the feature value of non-boundary points,divide the point cloud into feature subset and non-feature subset according to the feature value,and retain the points in the feature subsets.Finally,the improved farthest point sampling algorithm is used to simplify the non-feature subset.The algorithm can ensure the integrity of the model while deleting a large number of redundant points,and has a better retention effect on feature details.(2)Propose an adaptive point cloud simplification algorithm(APSA).Although the SPSA algorithm has a good simplification effect,the algorithm process is cumbersome and cannot be applied to batch processing tasks.APSA algorithm first calculates the feature value of each point based on FPFH,sorts and statistics the point cloud according to the feature value,uses the power-law function to calculate the probability value and cumulative probability value of each point,and then uniformly samples the cumulative probability value.The point corresponding to the cumulative probability value is the simplified point.Using the standardized information entropy as the evaluation standard,the experimental results show that APSA algorithm not only has higher entropy than other algorithms in small-scale point cloud model,but also has better simplification effect in large-scale point cloud model.(3)Constructs the cultural relic point cloud dataset Cultural Relic-3D.Since there is no public data set related to cultural relics,this paper collects the point cloud data of cultural relics from museum websites at home and abroad and obtains the point cloud data set of cultural relics after summarizing and sorting.(4)Propose a Multi-scale point cloud completion network(MPN).The network adopts the encoder-decoder structure,takes the multi-scale incomplete point cloud as the input,outputs the point cloud of the missing part,and optimizes the network with the GAN discriminator.MPN adopts the APSA algorithm to simplify the incomplete point cloud of the input point twice,and Cg MLP is used to extract the features of the incomplete point cloud of three different scales,and the features are merged and input into the MLP to obtain the feature vector.Eventually,point cloud data for missing parts is generated with the point pyramid decoder.The algorithm in this paper is compared with four related algorithms on the Shape Net and Cultural Relic-3D datasets.The results show that MPN algorithm has a good completion effect.