使用計算智能技術來辨識手寫中文字

Abstract

在一個圖型辨識的過程中大約可分為：一、前置處理，二、特徵擷取，三 、辨識比對等三個步驟。在這篇論文中，對這三個過程，我們均提出一種 處理方式。首先，在前置處理中，我們發展了一個切字的演算法，它可以 將一篇不工整文件內的字，獨立的切割出來。它具有良好的效率，並且已 移轉給廠商。接著，在特徵擷取方面，因為傳統的線性轉換方法在遭遇龐 大的維度時，極其困難，尤其是計算特徵值方面。因此我們提出利用類神 經網路的精神，找出一參考點，再利用此參考點來做特徵轉換，使比對的 特徵減少，以有效降低大量圖型比對的運算量。最後，應用此參考點方式 於 5401 個常用手寫中文字辨識過程的第一步驟，將字集規模大幅降低後 ，再細類分辨出字元，而完成整個辨識程序。 Generally speaking, the process of pattern recognition can be roughly divided into three stages, preprocessing, feature extraction, pattern matching. In this study, we try to accomplish the goal of hand-written Chinese character recognition. In each processing stage, we proposed new methods to cope with various problems. First, we propose an algorithm to cut characters out of a hand-written or printed document and separate them. It performs very well and has been transferred to manufacturers. Next, in the feature extraction stage, we propose a concept of reference vector. We apply reference vector(s) to transform the feature space from high dimension to low dimension. When the dimensionality is reduced, the time complexity of pattern matching will be decreased. Finally, we apply reference vectors together with other techniques for pattern matching.