Abstract
一個能夠妥善處理文字變形的方法是提高文字辨識系統效能的關鍵所在，我們稱這種變形的處理架構為文字變形模型(Character DeformationModel, CDM)，該模型是由文字的表示法和一套將文字變形加以模組化的機制所構成，在一般的辨識系統內部，文字變形模型所扮演的角色是參考字，而其中的輸入文字則被視為此模型的個案 (Instance)，我們根據這樣的基本想法設計用以建立高效能辨識系統所需的模型。在本論文中我們提出了兩種文字變形模型，一是限制型拋物線變換(ConstrainedParabola Transformation, CPT)模型，一是隨機式三次貝濟耳曲線(Stochastic Cubic Bezier Curve, SCBC) 模型，在限制型拋物線變換模型中的變換是用來描述文字變形的，參考曲線經由這些限制型拋物線變換作用後即視為對所觀察到之輸入文字的合理逼近，另一方面，在隨機式三次貝濟耳曲線模型中參考字是一連串的隨機式三次貝濟耳曲線和雜訊的組合，而輸入字便是其相對應之參考字隨機模型的量測值，這兩個模型都能精確地掌握住文字變形的特性，並作為建立高效能辨識系統的基礎。除了文字變形模型本身，論文中還提出了一個曲線的比對方法，稱作曲線調整程序(Curve Alignment Procedure)，用來建立參考字與輸入字之間的正確對應關係，整個程序分作兩個比對步驟，轉折點的彈性比對 (ElasticMatch of Turn Points)與合併-拆解過程 (Merge-and-Split Process)， 透過這個曲線調整程序，輸入文字曲線被分割成適當的小段曲線，而個別小段曲線和參考字模型的每一段隨機式三次貝濟耳曲線取得一對一的對應關係，這樣的對應關係可以保證獲得較可靠的文字距離度量以及較精確的參數估計結果，並且大幅度降低曲線比對時的計算成本。我們就所提出的方法進行了一連串實驗，根據實驗結果證實，不論是以限制型拋物線變換模型或隨機式三次貝濟耳曲線模型為基礎所建立的辨識方法都遠比早期的典型方法更為優異，在此，我們認為這種新的辨識方法已強化至足以有效應付草寫體文字的變形，而作為實用辨識系統的核心。本文共分為八章：第一章是概論；第二章說明階層式比對的觀念並提出曲線調整程序；第三章提出變換型文字變形模型；第四章提出隨機型文字變形模型；第五章對所提出模型的效能作進一步分析；第六章強化所提出的模型以適應個別的文字書寫方式；第七章建議模型的其他應用；第八章做一結論並指出未來的發展方向。A good method to manipulate the character deformations isessential to improving the performance of a recognition system.This manipulation framework is called a character deformationmodel (CDM). The CDM is composed of two components; one is acharacter representation and the other is a mechanism to modelthe deformations. In this framework, the CDM's serve as thereference patterns in a recognition system, and the inputcharacters are considered as the instances of these referencemodels. Based on the concept of manipulating the characterdeformations, we are devoted to designing the remarkable CDM'sfor the construction of a high performance recognition system.In this thesis, two CDM's, named constrained parabolatransformation (CPT) model and stochastic cubic B\'{e}ziercurve (SCBC) model, are proposed. In the CPT model, the inputcharacter is represented by the piecewise reference curves withthe CPT's that approximate the deformations. On the other hand,the reference character in the SCBC model is represented by asequence of SCBC's with some random noises, and the input oneis treated as the measurement of its related reference model.Both of the CDM's have the ability to accurately describe thecharacter deformations and are the foundation of constructing ahigh performance recognition system. Some experiments have beenperformed. As observed from these experimental results, therecognition methods based on the CPT and the SCBC models aremore effective and efficient than those proposed in sometypical previous works. It is believed that the proposedmethods are robust enough to act as the kernel of a practicalrecognition system.