面向拼接的文物碎块模型精细分类技术研究

There is a strong need for Computer-Aided automatic reassembling in broken cultural relics and protection work. However, current approaches when dealing with huge number of fragments face several challenges, including high model dimension, complex adjacency relationship, direct reassemble them is very hard, and etc. To cope with these challenges, this project proposes a two-stage reassembly solution, first classifying the fragments into parts and then reassembling them to reduce the number of reassemble pieces. According to the problem of little unlabeled of large number data and classification of high-dimensional data’s is not accurate, this project tries to improve classification accuracy using manifolds regularize, label propagation and semi-supervised theory to build classification model. And utilize the collaborative classifier theory to solve the classification problem of multiple types of high dimensional feature data. Based on the semi-supervised classification model, we try to reduce feature dimension and solve the problem of gaining the effective feature of classification algorithm to improve the usability. And the classifiable quantitative indicators is studied. We are planning to verify the efficiency and practicability of our solutions in real scenes by Qin Terra-Cotta Warriors fragments models. The scientific essence of this project is to discover inherent regular pattern of specific categories and low-level features of high dimensional model by adding local manifolds regularize and label propagation theory to semi-supervised classification, and discover essential connection between classification accuracy, robustness and model features of small labeled huge high dimensional data. The research findings of this project will provide theoretical support and application reference value for model classification and have a positive effort on intelligent classification and mosaics restoration of cultural relics.

在破碎文物拼接和保护工作中，迫切需要计算机辅助拼接修复，而面对大量碎块，现有的自动拼接方法却面临着模型维度高、邻接关系复杂、直接拼接困难等挑战。本项目提出先对碎块分类再拼接的思路减少拼接碎块数目；针对碎块已标注模型少和高维数据分类不准确的特点试图结合半监督分类、流形正则化和标签传递理论构建分类器模型，提高分类准确率；利用协同分类理论解决多种类型特征的模型分类问题；进一步研究半监督特征选择方法，解决有效特征获取问题；最后进行模型可分类量化指标研究。项目将通过秦俑碎块模型验证所提方法在现实场景下的有效性和可行性。其科学实质是在具有少量标注信息的大量高维数据中，引入局部流形正则化和标签传递理论进行半监督分类，寻求模型底层特征与具体类别的内在规律，探索分类准确率、鲁棒性和有效特征之间的本质联系。研究成果有望为模型分类提供有价值的理论支撑和应用参考，进一步为文物计算机智能分类和拼接复原产生积极影响。

在破碎文物拼接和保护工作中，面对大量碎块直接复原拼接难、工作量大、周期长等弊端，本项目提出先对碎块分类再拼接的方案以减少碎块拼接匹配工作量。主要针对模型复原中的特征提取、分类、分割和可分类量化指标、拼接四大关键问题展开深入研究：（1）针对非规则几何外形、受侵蚀而残损不全的碎块模型，项目提出了9个特征提取算法，提取了碎块模型几何形状、拓扑结构、统计信息、全局特征、显著特征、文物表面纹理特征、散乱点云特征和曲面类型分布等特征；（2）针对少量标注的高维数据精细分类问题，提出了5个分类算法，分别是：结合半监督流形假设的基于C均值聚类和图转导的半监督分类算法、半监督卷积神经网络特征提取和分类算法、基于迁移学习的半监督分类算法、基于残差学习的三维卷积神经网络模型、加入光滑性和一致性评价指标的多分类器协同训练算法；（3）研究了模型分割技术，提出了2个模型分割算法，然后将不同大小图像输入分类器，获得模型的可分类量化指标；（4）研究了碎块模型拼接修复问题，提出了9个拼接修复算法。这些都为破损兵马俑的虚拟拼接和复原过程提供指导和约束。. 本项目基于少量标注的高维数据精细分类问题，提出了多种特征提取、半监督分类和模型拼接方法，实现了碎块模型分类和拼接。共发表论文21篇，其中SCIE3篇/EI检索12篇；发明专利3项，其中授权1项；培养博士1名，硕士9名，远超5篇论文和2-3名学生等预期指标。. 课题研究运用模式识别方法，构建秦俑碎块模型分类和拼接体系，研究成果已应用于兵马俑碎块分类过程，可推广到其他领域，理论意义和应用效果显著。