應用統計遮罩比對方法與統計決策規則於古籍手寫文字之辨識

Using Statistical Mask Matching Method and Statistical Decision Rules to Recognizing Handwritten Characters in Chinese Paleography

我們提出了一個兩階段的分類方法，並佐以統計分析的技術，來辨識古籍手寫文字。在訓練系統時，我們使用網格碼轉換方法，把每個訓練樣本中最重要的離散餘弦轉換係數量化後，分配到一些網格中，使每一個網格代表一組具有相同量化值的文字類別；同時，我們也以統計方法為每個文字類別產生一個正遮罩和一個負遮罩。在辨識時，第一階段的粗分類先尋找與待辨識文字具有相同量化值的網格；第二階段的細分類則使用統計遮罩比對方法，將它與該網格中之所有文字類別的正、負遮罩逐一比對，再利用平均匹配機率之統計決策規則，來挑選相似度最高的類別做為辨識的結果。我們使用古籍金剛經來進行實驗，其結果証實了我們的方法非常適用於古籍手寫文字的辨識。

We present a two-stage classification approach, which is further enhanced by statistical techniques, to recognize regular handwritten characters in Chinese paleography. In the training phase, the grid code transformation method is applied such that each training sample is distributed to the related grid according to the grid code derived from its most significant DCT (discrete cosine transform) coefficients. Therefore, each grid consists of the character classes with the same grid code. On the other hand, for the purpose of fine classification, the statistical technique is also applied to generate a positive mask and a negative mask for each character class. To recognize an unknown character, the coarse classification is applied to obtain its grid code in the first stage. In the second stage, the fine classification that employs statistical mask matching is applied to match it against all the masks of those classes belonging to the grid. Then, the class with the highest similarity to the unknown character will be determined according to the proposed statistical decision rule based on average matching probability. We built an experimental system to recognize the Kin-Guan (金剛) bible. It is shown that our approach is effective in recognizing the handwritten characters of ancient calligraphers.