請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/65083
標題: | 手機拍攝手填表格辨識 Handwritten included Table Document Recognition in Mobile-captured Image |
作者: | Xue-Hwa Ang 洪婕華 |
指導教授: | 黃乾綱(Chien-Kang Huang) |
關鍵字: | 手寫,表格辨識,手機拍攝,文件分析,光學文字辨識OCR, handwritten,table recognition,camera-captured,document analysis,optical character recognition OCR, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 電子文件取代以往紙本資料的保存方式, 提高使用的便利性和即時性,是目前保存訊息的重要方式。隨着手機拍攝品質和電腦視覺的技術進步,手機攝影可代替傳統掃描。目前手機拍攝的文件掃描功能,雖然不能媲美掃描機的品質, 但手機拍攝不用額外準備掃描設備也可以即時使用,非常經濟和方便。在農場的環境限制,無法使用掃描設備拍攝,手機裝置即時拍攝手寫表格紀錄板 ,再進而人工輸入到數位電子檔。豬場環境狹隘和光線不足,平放對焦拍攝記錄板變得不易。因此拍攝的紀錄板的影像品質都一般影像差。紀錄板的文件都是表格結構,加上手寫文字,複雜的結構加上不規律的文字使得辨識變得困難。
目前手機掃描程式對於數位生成文件影像的辨識都有不錯的辨識結果,但對於複雜結構的文件和手寫的辨識都不太理想。手機拍攝影像除了會有嚴重的扭曲傾斜問題,有文件範圍以外的背景干擾,也有嚴重的光線不平均問題。有鑑於目前手機文件的對於表格手寫辨識文件技術門檻和需求,故本研究提出全新的三個階段辨識流程 : 第一階段,從手機拍攝影像(包含文件影像和非文件背景)中擷取出表格範圍的內容,第二階段從上步驟得到的表格影像範圍分別對篩選和處理得到表格的橫直方向的線邊緣,與原有邊緣合併,以封閉邊緣偵測格子。 本研究針對手填表格的格子偵測提出了新的評估標準。在本實驗收集的手機拍攝手填表格資料集,表格偵測的準確率爲93.33%,格子偵測的準確率爲99.67%。 As mobile devices equipped with high-performance digital cameras are widely available, demands to digitize printed materials with digital cameras have been rapidly increasing. It is because digital cameras have a number of advantages over the flatbed scanners, such as portability, fast response, and non- contact properties. In order to widen the areas of these camera-based applications, we address the table detection problem in the camera-captured images: This allows us to convert printed tables to computer-readable forms and we can use them in data retrieval, storing, and reproduction. The camera-captured document images will suffer from geometric distortions and uneven illuminations. Whereas the boundaries of tables are straight and parallel in scanned documents, they are neither straight nor parallel in camera-captured images. In the case of handwritten table documents, various artifacts complicate the task of get the accurate table’s line which is broken by overlapping with handwriting or color chip off. Also, when incorporated with the existing camera-based document analysis methods for the text-abundant images, this table detection method can improve the over- all layout analysis and optical character recognition (OCR) performances. So, table structure extract is necessary step before table’s document analysis. Our approach consists of three main stages: the first we do pre-process of camera-captured image to binary image which to localization the table boundary. Then, we use the angle properties to filter and recover the table’s line, and overlap with the line detects edge with original. Finally, we can detect the cell by using contours. With respect to our proposed method, we define a different criterion for handwritten included table’s cells detection evaluation. The precision of the table detection is 93.33% in and 99.67% for cells detection on mobile-captured handwritten included dataset. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/65083 |
DOI: | 10.6342/NTU202000539 |
全文授權: | 有償授權 |
顯示於系所單位: | 工程科學及海洋工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-109-1.pdf 目前未授權公開取用 | 4.57 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。