随着三维激光扫描技术的发展,点云数据已成为主流的三维模型格式之一.为解决大型佛龛和小型佛龛在相连情况下出现的错分现象,本文提出一种基于随机森林和多标签图割的石窟寺佛龛对象提取方法.以云冈石窟寺佛龛为研究对象,使用体素云连通性分割(voxel cloud connectivity segmentation,VCCS)算法对点云进行超体聚类,生成超体素块;通过递归特征消除法筛选得到最优特征集,并提取佛龛的点云数据;利用带噪声的基于密度的空间聚类(density-based spatial clustering of applications with noise,DBSCAN)算法和k 均值聚类,从佛龛点云中初步提取佛龛对象并赋予独立标签,通过多标签图割算法提取佛龛对象;并进行精度评价.结果表明:特征筛选前后分类精度有了明显的提升,总体精度由0.8762提升至0.9079;边缘的精细化提取效果得到了提升,提取总体精度达到0.9692.
A method for features extraction of Buddha Niches in Grotto Temples from point clouds based on random forest and multi-label graph cut
Buddha Niches hold a significant place in grotto temples and constitute an essential aspect of their digital preservation efforts.However,due to the detrimental effects of weathering and other afflictions,Buddha Niches have emerged as a focal point for conservation and restoration endeavors within these sacred sites.Addressing the challenge of low efficiency inherent in traditional manual extraction methods,coupled with the issue of misclassification between large and small interconnected Buddha Niches,necessitates the development of techniques for precise and efficient extraction of point cloud data pertaining to these Niches.To this end,this paper introduces a novel methodology for extracting Buddha Niche point cloud objects from grotto temples,leveraging the combined power of random forests and multi-label graph cuts.This approach aims to enhance the accuracy of object extraction for Buddha Niches within such temples.The proposed method encompasses several pivotal stages:(1)Point cloud preprocessing:This stage involves point cloud registration,denoising,and resampling,which collectively streamline the data complexity and boost subsequent processing efficiency.(2)Super-voxel block generation:Employing the voxel cloud connectivity segmentation(VCCS)algorithm,super-voxel clustering is performed on the preprocessed point cloud data,generating super-voxel blocks.This step significantly reduces the volume of point cloud data,enhancing classification and extraction efficiency.(3)Feature vector selection:Based on hypervoxels,a suite of descriptive features for Buddha Niches is devised,and the recursive feature elimination(RFE)method is utilized to identify the optimal feature subset for model training.(4)Random forest classification:The refined feature set is fed into a Random Forest classifier to achieve preliminary categorization of the Buddhist shrine point cloud data,thereby laying the groundwork for more intricate extraction processes.(5)Initial extraction of Buddha Niches:A hybrid approach combining density-based spatial clustering of applications with noise(DBSCAN)and k-means clustering is employed to provisionally extract Buddha Niches,assigning labels to individualized Buddha niche entities.(6)Refined object extraction:The initial extractions undergo optimization via a multi-label graph cut algorithm,enhancing extraction precision and disentangling neighboring Buddha Niches,ultimately ensuring accurate delineation.Experimental validation confirms the efficacy of the proposed point cloud processing technique,underscoring its superiority over conventional manual extraction methods and other standard algorithms in terms of both classification precision and edge detail extraction.Key outcomes include:(1)Enhanced classification accuracy:Feature screening elevated the overall classification accuracy from 0.8762 to 0.9079,demonstrating the algorithm's enhanced capability in discerning Buddha Niches.(2)Edge refinement efficacy:With an edge refinement extraction accuracy of 0.9692,the method exhibits exceptional prowess in managing Niches with indistinct boundaries.In summary,the presented methodology for extracting Buddha niche objects via random forests and multi-label graph cuts outperforms traditional manual extraction and conventional algorithms,particularly regarding classification precision and edge detail capture.Future research avenues will concentrate on further augmenting the automation accuracy for varying scales of Buddha Niches,refining feature selection mechanisms,and optimizing clustering algorithms to strike a balance between efficiency and high extraction fidelity.
3D laser scanninghypervoxelfeature screeningrandom forestDK clusteringmulti-label graph cuttingobject extraction
NETL
NSTL
万方数据
