文件傳達許多重要的資訊，如何將文件影像數位化並擷取文字訊息這個議題，隨著數位相機的普及而逐漸受到重視。為了獲得正確的文字辨識結果，一個以相機取像的中文文件辨識前處理系統，必須能處理不同排版格式、多種字體大小的行列、與使用者拍攝產生的文件輕微歪斜等問題，使得擷取出的文字區塊不會產生嚴重的謬誤。 中文文件與英文文件前處理最大的不同，是中文字由多個相連元件組成，如何將組成中文字的相連元件正確合併來進行中文字切割，是擷取中文字訊息最重要的步驟。本文提出的中文字的行列串連演算法與中文字部件合併判斷法則，可以克服影像小角度傾斜與中英字元混雜時的中文字合併問題，並能提供文字區塊讀序的資訊。另外論文中提出兩個文字訊息的保護機制:第一個是反白字區域偵測，通常反白字於相連元件抽取時會被視為背景雜訊過濾掉，為了保護資料完整性，有必要對反白字組成原件另外偵測；第二個是非正向文件的偵測，為了相機取像的便利性，以及使文件內容有較清晰的入鏡範圍，經常有將文件平面垂直光軸旋轉的需要，通常拍攝出的文件影像為接近正矩形的影像，但若要呈現正向的文字內容，可能仍需要旋轉0o、90o、180o、270o四種情形的校正。本研究提出一個以統計為基礎的方法，透過分析中文字筆劃的輪廓像素的方向性與文章中的中文字垂直投影波型，總合判斷中文文件的旋轉方向，提供文字識別模組，一個自動化的方向判斷機制。 本論文以名片測試本研究提出的辨識前處理系統，結果文字區塊正確切割擷取文字影像的成功率可達到98%，足以證明前處理系統設計方法的正確性。 As we know, Chinese documents convey a lot of meaningful and useful information. Due to the popularization of digital cameras, it is convenient to take picture and retrieve important text information from the digitalized Chinese document images. A successful camera-based Chinese document processing system should overcome the problems resulted from various document formats, font sizes, and document skewing to extract correct text block without generating erroneous results. The major difference between Chinese documents and English documents is that Chinese characters are mainly composed of multiple connected components. The most important step in obtaining the message of the existence of Chinese documents is to merge connected components with correct combining and produce complete Chinese character blocks. In this thesis, we propose a method to link Chinese characters into text line and develop a rule to discriminate the merging condition of ordering connected components to hypothesize the existence of skewing documents. Two mechanisms are developed in the thesis. The first mechanism is the detection of inversed text blocks which may be filtered out as oversize noise blocks in the preprocessing. The second mechanism is the detection of document images laid in incorrect direction because sometimes people will rotate camera 90o or 270o to capture document images. A two pass statistical method is proposed to automatically determine the rotating degree of documents images(0o、90o、180o、270o). The first step is devised by using the phenomenon that horizontal strokes appear more frequently than vertical strokes in Chinese characters. The second step is devised by analyzing the vertical projection histogram of each text block and defining keywords that assist in deciding the rotating degree.
