A Framework for Design Identification on Heritage Objects
A Framework for Design Identification on Heritage Objects
复制标题
DOI：
10.1145/3332186.3332190
复制
发表时间：
2019-07
期刊：
影响因子：
--
通讯作者：
Jun Zhou;Yuhang Lu;Karen Smith;Colin Wilder;Song Wang;P. Sagona;Ben Torkian
中科院分区：
文献类型：
--
作者：
Jun Zhou;Yuhang Lu;Karen Smith;Colin Wilder;Song Wang;P. Sagona;Ben Torkian
现代考古学中的一个具有挑战性的问题是，如何根据完整的装饰图案自动识别零碎的遗产物品，例如来自美国东南部的陶器碎片。这个问题的困难在于：1)这些陶片通常是零碎的，因此每个陶片只覆盖其底层完整设计的一小部分；2)这些碎片可能会高度退化，以至于曲线可能包含缺失的部分或变得非常浅；以及3)制作这些陶器时，曲线图案可能会相互重叠。本文提出了一个基于深度学习的框架，用于将SERD与已知设计的数据库进行匹配，以找到其底层设计。该框架包括三个步骤：1)使用基于FCN的曲线模式分割方法从数字化的SERD深度图中提取曲线模式；2)将SERD与非复合(设计的单一副本)模式匹配，并结合模板匹配算法和双源CNN重排序方法找到其底层设计；3)使用基于倒角匹配的方法将SERD与复合(设计的多副本)模式匹配。该框架是在一组来自划桨印花传统中心地带的碎片上进行评估的，这些碎片是北美殖民前东南部已知的划桨印花设计的子集。大量的实验结果表明了该框架和算法的有效性。
A challenging problem in modern archaeology is to automatically identify fragmented heritage objects by their decorative full designs, such as the pottery sherds from Southeastern America. The difficulties of this problem lie in: 1) these pottery sherds are usually fragmented so that each sherd only covers a small portion of its underlying full design; 2) these sherds can be so highly degraded that curves may contain missing segments or become very shallow; and 3) curve patterns may overlap with each other from the making of these potteries. This paper presents a deep-learning based framework for matching a sherd with a database of known designs to find its underlying design. This framework contains three steps: 1) extracting curve pattern using an FCN-based curve pattern segmentation method from the digitized sherd's depth map, 2) matching a sherd with a non-composite (single copy of a design) pattern combining template matching algorithm with a dual-source CNN re-ranking method to find its underlying design, and 3) matching a sherd with a composite (multiple copies of a design) pattern using a Chamfer Matching based method. The framework was evaluated on a set of sherds from the heartland of the paddle-stamping tradition with a subset of known paddle-stamped designs of Pre-colonial southeastern North America. Extensive experimental results show the effectiveness of the proposed framework and algorithms.