神經網路之階層式研究 : 應用於手寫中文字識別

Abstract

本論文主要在於研究如何應用階層式神經網路的分析技術做手寫中文辨識 。本論文之研究以從國小課本中選定六百零五個常用字為範圍，並以工研 院研定之手寫中文資料庫為訓練及測試樣本進行實驗，因測試資料不限定 個人或某團體，故所建立之系統具有廣泛性及一般性的辨識能力 。本系 統以傳統神經網路模式RCE及我們改良的INNRCE 作為學習比對之模式。系 統在掃描輸入文件之後，首先做了切字、非線性正規化、細線化等影像前 處理，接著抽取出二階特徵，二階特徵的優點是能表現筆劃的幾何特性， 因此作為辨識特徵可以提高識別率。因為中文字相當多，所以我們採用兩 層的辨識結構，在訓練階段，我們先求出學習集合中每字的中心，並訓練 好RCE 神經網路，在測試階段，取出文字的特徵後先與各類別中心求距離 ，選出前十個最接近的字作為候選字，這是第一層輸出，再以RCE 神經網 路進行辨識，這是第二層輸出。實驗結果在候選字階段平均分類正確率 為99.42%，整體識別率以我們改良的INNRCE 最佳，在不駁回時， 達91.23%,有97.10%落入前三名。我們並實驗了能辨識任意角度旋轉中文 的神經網路共三組，每組各十個字，結果以我們改良的INNRCE 最佳，在 不駁回時，平均識別率約為83％。 The purpose of this thesis is to apply hierarchical analysis techniques in recognizing handwritten Chinese characters. We select 605 most often used characters from primary school text books. The database we used comes from Industrial Technology Institute. Because the samples in this database are collected by more than 2600 people, our recognition system could reach a high generality and independence. We use a traditional RCE (Restricted Coulomb Energy) and a modified INNRCE (Incremental Nearest Neighbor RCE) as recognition model. Our recognition system includes preprocessing , feature extraction, candidate selection and word recognition. Since there are too many Chinese characters, we use two level recognition structure to reduce recognition time complexity. In training stage, we calculate each character's mean first and then train a RCE neural network for the 605 characters. In recognition stage, we calculate the distances between the input pattern and each character's mean and select the nearest 10 characters as candidates. These candidates are output of the first level. At the second level, we use RCE to recognize the designated character from the candidates. The candidate selection correction rate is 99.42% .Without rejection, the INNRCE has overall recognition rate about 91.23% and it reaches 97.10% within the first 3 candidates. In addition,we also construct three INNRCE neural networks to recognize rotational characters. Each network can recognize ten characters. Without rejection, the average recognition rate is about 83 %.