ON THE OCR APPROACH TO CREATING FULL-TEXT DATA BASE OF JAPANESE CLASSICAL LITERATURE

研究了一种新的消除干扰光学字符识别的图像噪声的方法。本研究的一个特点是利用颜色信息从红色字母、纸张、书写在半透明纸张背面的伪字母等图像噪声中提高“真”字母的分类能力，并以中国黑墨水书写在白色日本古典纸张上的日本原版古典书籍为研究样本，研究结果如下：（1）颜色分布特征：原始图像经彩色图像扫描仪（100 dpi，256灰度/R，G，B）数字化。并且每个图像单元被表示为RGB色度坐标中的三维矢量，然后被分析。颜色分布的特征是，（a）许多图像单元具有沿着R=G=B的线的颜色分布，（B）红色字母具有与（a）不同的颜色分布，（c）R、G和B颜色的亮度直方图几乎是双峰的。(2)图像的分类：（a）（1）中的（a）和（B）的特征对于区分红色字母和其他图像是有用的。(b)将判别阈值选择方法（Ohtu方法）应用于每个亮度直方图以确定黑色字母和纸段之间的阈值。该方法能很好地对两段文字进行分类，但容易切掉“真”黑色文字的外围图像单元。(c)通过引入聚类分析方法，对“真”黑色字母和纸段进行了更准确的分类，取得了较好的效果，验证了颜色信息对去除图像噪声的有效性。

A new approach to reducing image noises which disturb the optical character recognition has been studied. A peculiarity of the study is to use information about color to improve classification of "true" letters from image noises such as red letters, paper, pseudo-letters which are written on the reverse side of translucent papers and so on. Japanese original classical books written by the Chinese black ink on the white Japanese classical papers were selected as the research samples.The results are as follows :(1) Characteristics of Color Distribution : Original images were digitized by the color image scanner (100dpi, 256 gray-levels/R,G,B). and each picture cells are represented as 3-dimensional vector in the RGB-chromaticity coordinates then analyzed. The characteristics of the color distribution are, (a) many of the picture cells have the color distribution along with the line of R=G=B, (b)red letters have the different color distribution from (a), (c) brightness histograms of R,G and B colors are almost bimodal.(2) Classification of Images : (a) The characteristic of (a) and (b) in (1) are useful to distinguish red letters from another images. (b) The discriminant threshold selection method (Ohtu's method) was applied to each brightness histograms to determine thresholds between black letters and paper segments. This method can classify both segments sharply, but it is inclined to slices off the peripheral picture cells of the "true" black letters. (c) The cluster analysis was introduced to classify "true" black letters and paper segments more precisely, which gives better result.This study verify usefulness of the color information to eliminate image noise.