臺灣博碩士論文加值系統

本論文之目的是辨識文件中之單行文字。文件中的單行文字區塊取出之後，先作字元切割，再將切割完成後的字元送至分辨器作辨識。在字元切割方面，我們提出利用英數文字輪廓上下起伏的特性找出切割候選點的方法，並觀察英數文字的幾何特性以決定最佳的切割點組合。在字元辨識方面，我們提出綜合各種不同特徵以提昇辨識可信度的方式，先以Hidden Markov Model配合空間-時序特徵作候選者揀選，淘汰掉可能性較低的類別，縮小辨識範圍，接著將揀選出來的類別送至兩種整體特徵配合NNC作進一步的辨識，最後將三種分辨器的分數整合起來。實驗證明本研究之切割方法可有效切割多數字元（93.03%的正確率）；分辨 器對於數字之辨識率則可達95%以上（實驗包含兩種資料庫），對於NIST資料庫之英數字辨識率為73.44%，蒐集自高中男生與大學男女生的高中生手寫文字資料庫則為93.12%。對於由入出境簽證表單擷取出之文字區塊，其累積切割錯誤後之辨識率為79.62%。

This thesis addresses the problem of off-line hand-printed alphanumeric character recognition. Character separation procedure was first applied to break-off interconnected characters. The isolated characters are then classified into the best 30 out of the 62 possible classes using Hidden Markov Models. The final recognition was carried out by a nearest neighbor classifier using 64 features obtained from Kirsch operators and 33 contour features. The experiments were carried out using 12400 characters from NIST special database 19 and 6820 characters from our collections of hand-printed characters. The recognition rate of 95.4%/98.82% for numeric data and 73.44%/93.12% for alphanumeric characters were achieved for the two datasets, respectively.