Digital approaches to the capture and analysis of watermarks using the manuscripts of Isaac Newton as a test case

本研究将以牛顿的分散手稿语料库为例，探讨数位人文学术中两个普遍应用的研究领域。测试用例的直接目的是使用人工智能来帮助识别和分类牛顿材料中的水印，并在此过程中构建一个通用工具来帮助组织和确定手稿的日期。该项目还有更广泛的意义。该项目的第一阶段将是方法研究的技术，用于生产水印的图像，这是适合自动分析，使用新的摄影和探索潜在的现有图像。在第二阶段，我们将开发计算机视觉方法来系统地聚类和匹配手稿和收藏品中的水印图像集合。通过该项目开发的方法将可转移到牛顿语料库之外的水印收藏中，为寻求通过水印匹配分析，日期和组织历史收藏的学者创造了一种方法，并为寻求建立标准化测量和记录方法的保护者创建了一种方法，同时对水印文件进行成像和数字化。该项目的最后阶段将允许我们通过研究研讨会、网络工具和在线数据库的改进以及在期刊上的传统出版物来传播我们的发现。自20世纪初Charles Moïse Briquet的开创性研究以来，水印已经成为未注明日期的手稿的主要部分。Briquet在1907年出版的里程碑式的目录Les filigranes，原则上可以通过参考Briquet目录中的样本，确定研究人员在手稿中发现的1600年前的水印的日期（并在一定程度上定位）。由于伯恩斯坦协会（https：//memoryofpaper.eu/）的帮助，这一目录和其他目录已经数字化，但研究和技术的进步揭示了传统方法的局限性，这种方法需要耗时的程序和一定程度的专业知识来识别每一个水印。很难在原位水印和任何目录中复制的水印之间找到精确的匹配，首先是由于目录的全面性有限，其次是因为每个单独的水印都是以两个“孪生”版本生产的，从来没有完全相同，并且由于在造纸过程中的重复使用而随着时间的推移而变形。通过开发和增强新的方法和技术来改进水印的获取和分析，我们希望解决基本问题，从而为所有必须依赖纸质文档获取时间顺序证据的人提供好处。由于机器学习和人工智能，近年来计算机视觉取得了重大进展，该项目将建立在国家图表学院及其合作伙伴（特别是巴黎蓬茨理工学院的计算机科学家）已经开展的尖端工作的基础上，以研究跨格式（照片和跟踪）匹配图像，特别是水印的问题。在创建用于训练和开发由Ecole des Chartes创建的开源软件的图像语料库时，我们将以国家档案馆（TNA）最近的工作为基础，使用相对负担得起的设备和技术来生成非常适合机器分析的水印图像。该项目将开发和应用这两种方法，以尝试增强计算机视觉软件，使其能够解锁数千张反射光拍摄的现有图像中的潜在信息，这些图像已被机构数字化并通过IIIF访问。

This project will investigate two research areas with general application in digital humanities scholarship, using the dispersed manuscript corpus of Isaac Newton as a test case. The immediate purpose of the test case will be to use artificial intelligence to assist with the identification and classification of watermarks in Newton material and, in the process, to build a general tool to assist with the organisation and dating of manuscripts. The project also has much wider significance. The project's first stage will be the methodological investigation of techniques for the production of images of watermarks which are suitable for automated analysis, using both new photography and the exploration of the potential latent in existing images. During the second stage, we will develop computer vision methods to systematically cluster and match the assembled corpus of watermark images across manuscripts and collections. Methods developed through this project will be transferrable to watermark collections beyond that of Newton's corpus, creating a methodology for scholars seeking to analyse, date, and organise historical collections via watermark matching, and for conservators seeking to establish standardised surveying and documentation methods while imaging and digitising watermarked documents. A final stage of the project will allow us to disseminate our findings through research workshops, web tools, and improvements to online databases, as well as traditional publications in journals.Since the groundbreaking early twentieth-century research of Charles Moïse Briquet, watermarks have formed a central part in the dating of otherwise undated manuscripts. Briquet's monumental 1907 catalogue, Les filigranes, made it possible, in principle, to date (and to some extent localise) pre-1600 watermarks found by researchers in manuscripts by reference to exemplars in Briquet's catalogue. While this catalogue and others have been digitised thanks to the Bernstein consortium (https://memoryofpaper.eu/), advances in research and technology have revealed the limitations of the traditional approach, which requires time-consuming procedures and some degree of expertise for the identification of each single watermark. It is very difficult to find exact matches between watermarks in situ and those reproduced in any catalogue, first due to the limited comprehensiveness of the catalogues, and, second, because each individual watermark is produced in two "twin" versions, never perfectly identical, and suffers deformation over time as a result of repeated use in the paper manufacturing process. By developing and enhancing new approaches and techniques to improve the acquisition and analysis of watermarks, we hope to solve basic problems and thereby provide benefit to all who must rely upon paper documents for chronological evidence. While computer vision has made significant progress in recent years thanks to machine learning and artificial intelligence, this project will build on cutting-edge work already undertaken by the Ecole Nationale des Chartes and its partners (notably the computer scientists at École des Ponts ParisTech) to investigate the problem of matching images, specifically of watermarks, across formats (photographs and tracings). In creating a corpus of images used to train and develop the open source software created by the Ecole des Chartes we will build on recent work by The National Archives (TNA) to use comparatively affordable equipment and techniques to produce images of watermarks that are highly suitable for machine analysis. The project will develop and apply both of these approaches in order to attempt to enhance the computer-vision software so that it may be able to unlock the latent information held in thousands of existing images shot in reflected light which institutions have already digitised and made accessible through IIIF.