A Siamese Network-Based Method for Automatic Stitching of Artifact Fragments
A Siamese Network-Based Method for Automatic Stitching of Artifact Fragments
复制标题
DOI：
10.1109/tim.2023.3295018
复制
发表时间：
2023
影响因子：
5.6
通讯作者：
Quanjiang Liang;Li Yang;Zai Luo;Wensong Jiang;Chengkang Hong
中科院分区：
文献类型：
--
作者：
Quanjiang Liang;Li Yang;Zai Luo;Wensong Jiang;Chengkang Hong
针对文物碎片由于自然或人为因素造成的碎片化问题，提出了一种基于连体网络的文物碎片自动拼接方法。首先，该方法采用了一种改进的区域增长分割算法分割的断裂和非断裂表面的点云的工件碎片。其次，使用刚体力学模拟方法对虚拟工件进行碎片化，并建立碎片数据库，用于深度学习算法训练。然后，利用神经网络动态图CNN（DGCNN）-Siamese网进行点云相似性比较，实现破碎件断口的匹配。第三，通过Harris三维特征点提取、邻域点特征直方图（PFH）特征描述和迭代最近点（ICP）方法对断口点云进行配准。实验结果表明，该方法的整体匹配准确率为96.99%，通过对比分析，该方法能够减小配准偏差，实现碎片伪影的完全恢复。
For the problem of fragmentation of cultural relic fragments caused by natural or man-made factors, this article proposes a method of automatic splicing of cultural relic fragments based on the Siamese network. First, the method employs an improved region-growing segmentation algorithm to segment the fractured and nonfractured surfaces of the point cloud of artifact fragments. Second, a rigid-body mechanics simulation method is used to fragment virtual artifacts and establish a database of fragments for deep learning algorithm training. Then, point cloud similarity comparison using a neural network dynamic graph CNN (DGCNN)-Siamese net to achieve matching of fracture surfaces of broken pieces. Third, the fracture surface point cloud registration is aligned by using Harris-3-D feature point extraction, neighborhood point feature histogram (PFH) feature description, and iterative closest point (ICP) method. The experimental result shows that the overall matching accuracy of the method is 96.99%; the method is able to reduce the registration deviation and achieve complete recovery of the fragmented artifacts through comparative analysis.