連筆手寫數字辨識系統之研究

Abstract

本論文的目的是設計一套能切割連筆數字的辨識系統。連筆數字切割一直 是數字辨識的核心問題，其困難的原因是由於數字之間相連的部份和數字 的正常筆劃部份差異不大，因此要從眾多筆劃之間決定正確的切割點是件 不容易的工作。而且數字相連和破碎的情形皆有可能同時發生，再加上手 寫字的變異度大，所以要設計一個完整且具高效能的手寫文字辨識系統， 被公認為是一個困難且極具挑戰性的問題。本論文提出一個新的理念：要 解決連筆數字辨識問題除了切割之外，需再加上理解與辨識的功能，兩者 交互配合，方能求得最佳的切割點。我們設計的辨識系統由三個部份組成 ，分別是（一）切割部份、（二）理解部份以及（三）辨識部份。系統的 整體辨識率達 80.67%。本論文所有的測試及訓練樣本皆來自美國郵政總 局所出版的手寫字資料庫，無論是測試的結果以及所建立的系統都頗具客 觀性和一般性。 This Thesis develops segmentation techniques for touching numeral strings.The segmentation of touching numeral strings has been known difficulty problems, due to the simularity between the touching point and the join point of normal numerals.In addition, touching and broken strokes as well as distortion characters make the proper segmentation even more difficulty.In this research, we propose a new concept for this matter: proper segmentation needs to be interactive with numeral reconition process. Therefore, we develop a system contains three parts: (1) segmentation subsystem, (2) understanding subsystems, and (3) recognition subsystem. The segmentation subsystem performs touching detecting and cutting processes. A neural network and an expert system are designed to understand and to recognize segmented numeral string. Based on our experience we are able to cut and to recognize touching numeral string with 80.67% of correct rate. In order to be compatible and compitible with other system, we use the numeral data base published by CEDAR of SUNY for either system training and testing.