固定表格手寫中文之切割

Abstract

固定表格手寫中文字之切割與辨識研究生：許文瑞 指導教授：李 錫堅博士 國立交通大學資訊工程研究所 摘 要 本論文是介 紹在固定表格手寫中文字切割的方法。在論文的第一個部份，我們將介紹 三個關於手寫中文字切割的方法，分別是投影(projection operation)、 相連元件抽取(connected component extraction)和最短路徑切割(short path cutting)。為了要有效率的抽字，我們使用水平和垂直投影來抽取 文件上的中文字。第一次的水平投影是用來決定一個欄位有多少列的字， 如果在最欄位的最高級最低的現有超過一定量的點數，我們就會考慮欄位 外的黑點。接著的垂直投影是為了將字分開，如果有兩個元件比較近，我 們就將之合成一個字。最後的水平投影是為了找出字超出格現外的部份。 因為格線的座標已知，我們直接將格線去除。然而有些被切出的字組其寬 度比平均寬度大，我們使用相連元件來抽取。而有些字可能還連在一起， 我們便使用最短路竟的方法來切割比平均寬度的1.5倍大的字組區塊。 在論文的第二部份，統計式的辨認模組來辨識所有從文件中抽出的中文字 。因為數字的寬度比較小，所以會被合成一個中文字，當這些誤抽的區塊 輸入時，辨認模組將會根據他們的差異性（difference） 而拒認。我們 系統的抽出率可達89.40%而系統的辨識率則是55.23%。 In this thesis, methods for handwritten Chinese character segmentation in form documents with known structure are introduced. In the first part of this thesis, three steps are proposed to extract handwritten characters, named projection operations, connected component extraction and shortest path cutting. In order to extract the character efficient, we perform vertical projection and horizontal projection. The first horizontal projection is used to determine the number of text lines in a grid. If a sufficient number of points of a text line exceeds the top or bottom grid line, the boundary rectangles of text lines can be extracted outside the grid. The next vertical projection is used to segment the characters in a text line. If two small components are close, they are merged into a character. The last horizontal projection tries to find the top and bottom boundaries of certain characters outside the grid. Because the coordinates of all form lines are predefined in the database, the form lines can be removed from characters. Since some very close characters may be segmented into a character block whose size is greater than the average size of character blocks, we next perform the connected component analysis in the blurring character block. Third, since some Chinese characters may still touch each other, we use shortest path cutting to separate the character block whose width is greater than one and half average width. In the second part of this thesis, characters extracted from the form document are sent to a statistic character recognition module. Some numerals may be grouped into a large character block because their widths are smaller than characters widths. These mis-merged characters will be rejected by the recognition module due to their large differences. The extraction rate of our system is up to 89.40%, while the recognition rate of our system is 55.23%.