印刷不良名片英文與數字之辨識

Abstract

在這篇論文裡我們設計一種方法用來辨識名片上的英文與數字。此方法包括三個部分，即文字辨識核心、連字切割與破碎字合併模組和辨識結果修正模組。首先，我們將單行的二值化影像做水平校正、斜體字偵測、連通單元(connected components)抽取、適當的合併與去雜訊、斜體字推正與平滑化。將框住連通單元的矩形由左向右水平投影，找出最靠近上下兩側的最大投影量，決定一行文字中的四條基準線。抽取出來的連通單元，由一個統計式方法設計的辨識核心程式辨識。我們用三種判斷連字的特徵，即寬高比(aspect ratio)，辨識結果的可靠度(recognition cost)及水平方向的筆畫變異度(horizontal crossing-count)，以區別單一文字與連字。接著以兩階段的切字模組進行切字。第一階段以分析連字影像的垂直投影決定可能的切點，如果第一階段中的切點導致切字的失敗，第二階段再分析連字的上下輪廓決定其他可能的切點。我們試著將相鄰且靠得很近,同時又是較細的字元(aspect ratio小)或碎點(例如:逗號和句號)合併，分析合併後字元之辨識結果，將可能的破碎字(broken characters)辨識出來。在雙語言的辨識系統中，例如:中英文名片辨識系統，我們討論三種信心度(confidence values)用來決定連通單元是中文或英數字，再分別以不同的辨識核心辨識。最後，根據連通單元的垂直位置和字元高度、一行文字的四條基準線、前後文的資訊與字元影像結構分析，調整辨識結果，使得辨識核心容易混淆的字元也能辨識正確。 我們從300張英文名片上擷取約10,000個英數字影像，作為測試辨識核心的樣本，測試的結果，辨識率達到了99.47%；我們擷取479個相連文字，其中有92.90%被正確地切割出來; 另外，我們也擷取336行的單行文字，找四條基準線的正確率達到了98.51%。

In this thesis, we design a procedure for recognizing characters in name cards. The procedure consists of three main operations: character recognition kernel, character segmentation, and correction of recognition results. In the first phase, we get possible isolated characters by performing the pre-processing, which includes binarization, skewing angle detection and straighten, italicness detection, smoothing, and connected-components extraction. We project the bounding rectangles of connected-components from left to right on the horizontal direction. Four typographical lines of single text lines are determined by projection analysis. Then, the components are recognized by a statistical multi-font recognition kernel. To detect touched characters, we use three measurements include aspect ratio, recognition cost and horizontal crossing-count. In the character segmentation phase, we develop a two-stage character segmentation module to segment the touched characters. In the first stage, we find the breaking points by projection analysis. If all possible segmentations have bad recognition results in the first stage, we use outline analysis to find other breaking points in the second stage. In order to solve the problems of broken characters, the merging process are give in our system. We try to merge consecutive and close characters that are either thin characters (small aspect ratio) or fragments (such as "," and "."). We analysis the recognition results of merging characters to decide whether accept or not. In some dual-lingual systems, such as Chinese-English name card understanding systems, we propose three measurements of confidence values that determine which language connected-components belong to. Then, either the Chinese OCR kernel or the English OCR kernel will recognize connected-components. Finally, the geometric properties, contextual information and disambiguation by structural difference are used to correct the recognition results. By rejecting impossibilities, the correct classes may eventually be promoted to the first candidate. Moreover, characters that have the same shape in capital and lower cases are justified according to their geometric properties. In our experiments, we extracted about 10,000 single characters and 479 touched characters from more than 300 English business name cards as test samples. Besides, we also collected 336 images of text lines to test the four typographical lines determination module. The recognition rate for single characters was 99.47%. 92.90% of touched characters were correctly segmented. The accuracy of determining four typographical lines was 98.51%.