古籍影像與文本之對應－以《古今圖書集成》為例

Abstract

《古今圖書集成》為現存最大類書，因此有不少數位人文學者將其與資料庫系統結合，做成《古今圖書集成》全文檢索系統，內容大多包含文字及影像的搜索功能，但在結果的呈現上皆重於文字，對影像的部分並無多加著墨，所以當使用者想從影像中獲取一些資訊，例如找某個關鍵字詞時，只能用肉眼觀察影像的內容，無法從系統提供幫助。 在本研究中，試圖避開OCR技術的輔助，直接對影像及文本處理，讓兩者間有高度的對應關係，再利用文本來尋找文字在影像中的位置。首先對所有影像做一些影像處理，包含了旋轉與切割，使每張影像有著相同的格式與排版，再分析影像特性，如：文字的排版方式、影像中圖像有固定大小與位置等等，利用這些特性以行為單位將影像的狀態完整對應到文本中，最後文本每一行對應到影像中文字、空行、圖像三種狀態其一。 最後再利用對應完成的文本及處理過的影像，先計算文字在文本中的位置，再透過對應座標的方式找出文字在影像中的位置。如此使得《古今圖書集成》影像將不再只是以插圖的形式點綴系統，而是能實際提供有用的資訊給使用者。

The Complete Collection of Graphs and Writings of Ancient and Modern Times (Gujintushujicheng, or Jicheng for short), completed in the early 18th century, is the largest book in the world in existence. Containing over one million Chinese characters, almost 100,000 pages, and cover over 6,000 subjects, Jicheng is also difficult to use. During the past decade, several digital systems have been developed so that people can use Jicheng through fulltext search. However, all of these system did not attempt to match images and texts, which would make using Jicheng even easier. This difficult arises partly because for old Chinese books, OCR is still not an effective technology. In this thesis we develop a method that tries to find direct correspondence between an image of Jicheng and its associated text without resorting to OCR. We first calibrate the images so that all 100,000 pages in the book have the same size and format. We then analyze the characteristics such as the format, number of lines, position of graphs, etc, so that each line in the typed text maps to either a line of text, a blank line, of part of a graph in a page image. Once this is done, we then do a character-by-character mapping between each character in the typed text and a character in a page image. Our method is quite effective. The accuracy in mapping the entire contain of Jicheng is 98,7%. The rest is mainly due to typographic errors occurred when typing the full text, which can be easily corrected by hand.