本文主要探討字元的不變性辨識,提出一種神經網路的方法來有效解決字元辨識時位移、比例變化或旋轉等情形的發生,使其都能有效地將相同字元認出並歸屬於同一類。為了達到字元的不變性辨識,本文採用了「文字本體的像素至文字重心的距離變化量」為主要特徵,此特徵與字元的位移、比例大小及旋轉角度無關;利用影像特徵處理,先擷取已定義的特徵值,然後進行分群處理,後再用SimNet神經網路學習辨識的能力,將這些特微值予以分類辨識。而本文依提出的方法建構一套系統,來驗證所提出之方法,在實驗的結果中,單獨針對位移情形的辨識率達100%,單獨針對比例變化情的辨識率達99.6%,單獨針對旋轉情形的辨識率達97.2%。而在針對位移、比例變化及旋轉交互發生情形的整體辨識遇程中其辨識率達92%,故無論影像產生了位移、比例變化及旋轉情形都可以使用本文提出的方法加以辨識,可以有效地處理字元的不變性辨識。
Abstract This thesis proposes a neural network approach to recognize the invariance of characters including Chinese and English letters. The character invariance recognition takes position, scale and rotation into account. To find the features invariant to a character, the statistical distance estimates between the pixels and the gravity of the recognizing character have been developed. The preprocessor takes and classifieds those invariant features that are recognized by the proposed neural system SimNet. After the experimental recognition, the accurate rates 100%, 99%, and 97% for position, scale and rotation, respectively, were found. The rate of the recognized accuracy for the variety of the invariance averages 92%. The high accurate recognition shows that SimNet is a potential character recognizer.