以階層式類神經網路作手寫中文字辨識

Abstract

本論文提出一個以階層式的類神經網路所組成的手寫中文字辨識系統。本 系統由兩個子系統相連接構成：大分類子系統和細分類子系統。大分類子 系統是由多層的柯荷能(Kohonen)網路組成，並執行改良的贏者通吃( winner-take-all) 學習演算法。首先，將字元影像正規化為相同大小的 字；接著，再利用細線化模組將正規化後的影像轉化為字元骨架；之後， 分析字元骨架中的各影像點屬於四個筆劃方向平面的可能程度，以作為大 分類子系統辨識基礎。細分類子系統主要是由下列幾個模組所組成，分別 為筆劃分向模組、筆劃預比對模組、反覆式比對模組、以及相似度量度模 組。本研究提出一個自動選取參考字集的方法。在大分類階段，採 用1,000 個字集、每個字集取12 個樣本做為類神經網路訓練之用，我們 將它分類成 130 群，再利用CCL-HCI1 10,000筆資料做測試，可以達到 94%的平均大分類正確率；而細分類方面，共建立3 套子系統，可辨識 143 字，達到 92.2% 的平均辨識率 A hierarchical neural network system for recognition of handwritten Chinese characters is proposed. The system is composed of two subsystems: the preclassification subsystem and the detailed matching subsystem.The pre- classification subsystem includes multiple Kohonen networks which perform a modified winner-take-all learning algorithm. At first, a character image is normalized. A thinning module converts the normalized image into a skeleton. The probability with which each image point of the skeleton belongs to a directional stroke plane is computed for preclassification. The detailed matching subsystem is composed of several modules ,namely, the line separation module, the stroke prematching module, the iterative matching module, and the similarity measuring module. A voting approach to decision making using five reference character sets is employed. A method for automatic selection of reference character sets is also proposed. 1,000 character classes and 12 characters randomly chosen from each class were used as training samples. The totally 12,000 characters are classified into 130 clusters. The preclassification subsystem was tested by 10,000 other characters to yield a 94% average rate of correct classifications. Detailed matching subsystems for testing three clusters of 143 characters were built and a 92.2% average recognition rate has been achieved.