A Duplicate Chinese Document Image Retrieval System

Introduction

An optical character recognition (OCR) system enables a user to feed an article directly into an electronic computer file and translate the optically scanned bitmaps of text characters into machine-readable codes; that is, ASCII, Chinese GB, as well as Big5 codes, and then edits it by using a word processor. OCR is hence being employed by libraries to digitize and preserve their holdings. Billions of letters are sorted every day by OCR machines, which can considerably speed up mail delivery.

The techniques of OCR can be divided into two approaches: template matching and structure analysis (Mori, Suen & Yamamoto, 1992). The template matching approach is to reduce the complexity of matching by projecting from two-dimensional information onto one; the structure analysis approach is to analyze the variation of shapes of characters. The template matching approach is only suitable for recognizing printed characters; however, the structure analysis approach can be applied to recognize handwritten characters.

Several OCR techniques have been proposed, based on statistical, matching, transform and shape features (Abdelazim & Hashish, 1989; Papamarkos, Spilioties & Zoumadakis, 1994). Recently, integrated OCR systems have been proposed, and they take advantage of specific character-driven hardware implementations (Pereira & Bourbakis, 1995). OCR generally involves four discrete processes (Khoubyari & Hull, 1996; Liu, Tang & Suen, 1997; Wang, Fan & Wu, 1997):

• 1.

  separate the text and the image blocks; then finds columns, paragraphs, text lines, words, and characters;

• 2.

  extract the features of characters, and compare their features with a set of rules that can distinguish each character/font from others;

• 3.

  correct the incorrect words by using spell checking tools; and

• 4.

  translate each symbol into a machine-readable code.

The duplicate document image retrieval (DDIR) system transforms document formatted data into document images, then stores these images and their corresponding features in a database for the purpose of data backup. The document images are called duplicate document images. When retrieving a duplicate document image from the database, users input the first several text lines of the original document into the system to create a query document image. Then the system figures out the features of the image, and transmits to the users the duplicate document image whose image features are similar to those of the query document image (Nagy & Xu, 1997).

Some approaches have been proposed for the DDIR system. Doermann, Li, and Kia (1997) classified and encoded character types according to the condition that four base lines cross each text line, and uses the codes as the feature of the document image. Caprari (2000) extracted a small region from one document, assigned this region to the template (signature generation), and then scanned this template over a search area in another document. If the template also appears in the second document (signature matching), the two documents are classified as duplicates. Angelina, Yasser, and Essam (2000) transformed a scanned form into a frameset composed of a number of cells. The maximal grid encompassing all of the horizontal and vertical lines in the form is generated; meanwhile, the number of cells in the frameset, where each cell was created by the maximal grid, was calculated. Additionally, an algorithm for similarity matching of document framesets based on their grid representations is proposed too. Peng, Long, Chi, and Siu (2001) used the size of each component block containing a paragraph text image in a duplicate document image and its relative location as the features of the duplicate document image.