手寫中文字辨認之決策型類神經網路研究

Abstract

本論文的目的是應用類神經網路於手寫中文文字之辨識。本論文所能辨識 的字為教育部所選定的 5401 個常用字，並以工研院研訂之手寫中文字資 料庫為訓練及測試樣本進行實驗，因為測試資料不限定於個人或某團體， 故所建立之系統具有廣泛性及一般性。本辨識系統包括四個步驟，依序是 前處理、特徵擷取、大分類以及細分類。在前處理的部分，我們經分析比 較，採用了最佳之影像處理流程，包括平滑化、大小正規化、非線性正規 化以及細線化。在特徵擷取流程，我們藉由一特徵指標來選取優良的文字 特徵，使特徵擷取流程和辨識流程能獨立進行。在大分類辨識流程，我們 採用了重疊式的群集演算法作為大分類器，以加快系統之辨識速度。在細 分類辨識流程，我們使用決策型貝氏神經網路作為細分類器，進行單字之 辨識。實驗結果將 5401 個常用字分成 60 個大分類，其正確率為 99.9%以上；整體之辨識率為 86.68%，而前三大候選字正確率為 93.60%，前十大候選字之正確為 95.94%。 This thesis presents an application of neural networks on off- line handwritten Chinese characters recognition. Our recognitin system includes preprocessing, feature extraction, coarse classification and fine classification. There are four steps in preprocessing stage, they are smoothing, linear normalization, nonlinear normalization and thinning. In the stage of feature extraction, we proposed a feature index that indicates the quality of various features. On the recognition stage, we propose a two-level recognition structure in our system to reducere the recognition time complexity greatly. In the coarse classification stage, we use overlapped cmean clustering algorithm to implement a coarse classifier, which can successful reduce about four fifth of computation time. In the fine classification, we use Bayes decision based neural network as our fine classifier. In order to evaluate the proposed recognition system, we choose 5401 frequently used Chinese characters as our trainning and testing domain. The database of each testing and trainning sample character was provided by the CCL of ITRI. Because the samples in this database were collected from more than 2600 persons, our recognition system could reach a high generality and user-independence. Experimental results show that, for the outside testing the recognition rate is about 86.68%, and for the TOP3 recognition, it reachs 93.60%.