中文文字識別(Chinese optical character recognition)技術在近二十幾年內，因為許多研究者的投入努力而有長足的進步，甚至已成熟至商品化的階段，這樣的發展結果說明了兩個事實：中文字文字識別的需求及在技術上的可行性。文字識別在各種領域中的廣泛應用，從早期的銀行支票處理及郵件中的郵遞區號自動分信，到辦公室自動化的文件數位化，以至近來知識管理及數位圖書館的知識檢索(information retrieval)，無不日益仰賴文字識別的技術。 本論文的主要目的有二：一、以單一樣版字型資料庫做多字形的中文字分類。二、對旋轉字做識別。而為了適應新的文字識別需求，使用較少的系統資源及較快速的辨識處理時間是研究的預設限制。 我們提出了一個全新的形似骨架產生法，並以此虛擬骨架為基礎，擷取文字的特徵以做為辨識的依據。在論文中我們討論了三種不同形態的虛擬骨架及對應的虛擬輪廓，在前處理時間的節省上有極大的改善。而以虛擬骨架為基礎，將二維的文字影像以一維的字碼特徵表示，是在特徵擷取階段透過對虛擬骨架的投影梯度圖，產生文字的假筆劃字碼，以做為分類階段的依據。在分類階段，我們以變異權重的編輯距離演算法及模糊編輯距離演算法對文字的特徵字碼進行分類。以五千四百零一個常用中文字為測試樣本，我們以最常用的細明體為樣版字集，再分別以不同大小的標楷體及細明體字，驗證我們所提出方法的成效，結果顯示，在時間及記憶體空間的需求上，都可以大幅改善文字分類的效能。 另外針對旋轉文字的辨識，論文中以文字的虛擬輪廓為基礎，以模糊C-mean演算法及我們所提出的循環模糊C-mean演算法，擷取旋轉文字的特徵向量，再以循環漢明距離估測文字間的相似度，以做為文字辨識的依據。在實證上我們以象棋中的十四個棋子文字為測試樣本，以四種不同旋轉角度驗證而得所提出方法的正確及可行性。 In this dissertation, a novel method is presented to classify machine printed Chinese characters by matching the code-string-based features which are generated from pseudo skeleton. In our approach, the proposed novel pseudo skeletons of Chinese characters are extracted instead of the skeletons generated by the traditional thinning algorithms. The features of the pseudo skeletons of input and template characters are encoded into two code strings. Next, the edit-distance based matching algorithm is employed to compute the similarity of two characters based on their corresponding encoded strings. There are three main modules in our work which include preprocessing, feature extraction, and fuzzy matching modules. First, p-skeletons of an input character and the pixel projection histograms are generated in the preprocessing module. Three kinds of virtual-strokes (called v-strokes) are defined by using the fuzzy membership functions. These features are encoded and represented by three kinds of fuzzy variables in the feature extraction module. Based on the encoded strings, the problem of OCR classification is transformed to the matching problem of 1-D string instead of that of 2-D image. At the training stage, the extracted features are stored in the reference database, whereas the fuzzy edit-distance matching algorithm is applied to measure the similarity of an unknown pattern and those in the reference database at the classification stage. Finally, the candidate list is generated as the classification results. Experiments were conducted on 5401 daily-used Chinese characters of various fonts and sizes. Experimental results are illustrated to demonstrate the validity and efficiency of our proposed method. The main contribution of this dissertation is to effectively classify the multi-font Chinese characters using single-font reference database. In addition, a new method for rotational character classification is also proposed in this dissertation. Similar to p-skeleton generation, the pseudo contour of a character is generated first. The using of pseudo contour instead of original image can greatly speed up the process time. A new clustering method, called circular fuzzy C-mean algorithm, is devised to obtain the rotation invariant feature. At the classification stage, the Hamming distance is applied to measure the similarity of the characters. Experiments shown that the fuzzy ring feature is effective for rotational character classification.
