手寫辨識可應用在很多不同的領域。例如驗證簽名的真假,辨識信封上手寫的地址和區碼,辨識在PDA 或 TablePC上手寫的文字等。在這篇論文中,我們只要研究獨立文字辨識,主要是因為它是很多複雜的手寫辨識系統中的基本元件,改善它的準確性有助於提升整個系統的準確性。我們會比較四個不同的方法,包括主要成分分析,線性鑑別分析,獨立成分分析和非負矩陣分解。結果顯示主要成分分析準確率最高;線性鑑別分析準確率最低;獨立成分分析跟主要成分分析效能差異不大;當考慮類別資訊時,非負矩陣分解在每個類別中只需很少的基底圖像就可達到不錯的效果。
There are many applications for handwritten character recognition, such as signature verification, handwritten address recognition, pen-based input method used in PDA etc. In this thesis, we just consider offline character recognition because it is the basic building block of many complicate handwriting recognition system. We compare four techniques for handwritten recognition. They are PCA, LDA, NMF and ICA. The result shows that PCA has the highest accuracy. LDA has the lowest accuracy due to small training data set. The difference of performance between ICA and PCA is small. NMF only need smaller number of basis images within each class when considering class information.