本論文提出在中文字辨識領域中以鄰近搜尋的方式取代深度學習的模型訓練架構。採用以Two Stage方式，將教育部提供的4808個常用中文字作為文字辨識的依據，以影像形態學處理加上水平垂直投影的方法進行文字切割，使用灰度共生矩陣與空間矩擷取中文字的特徵，透過正規化將特徵值等比例的縮放到0~1區間輸出，並以不同字型樣式的4808中文字作為優勢點樹分類器資料庫，並以優勢點樹分類器透過歐幾里得距離範圍進行中文字的鄰近搜尋辨識，與開源的Tesseract-OCR光學字元辨識軟體進行4808個中文字常用的辨識結果比較。實驗中發現優勢點樹分類器的建立時間均低於1秒，比起深度學習模型的訓練減少許多，而且在以新細明體作為優勢點樹分類器資料庫對於不同字型中文字的鄰近搜尋中，平均辨識率達到79%，優於Tesseract-OCR中文字的辨識結果。;This paper proposes a model training architecture that replaces deep learning with proximity search in the field of Chinese character recognition, use 4808 commonly used Chinese characters provided by the Ministry of Education as the basis for character recognition, by using image morphology processing plus horizontal and vertical projection for Chinese character cutting, Using the grayscale symbiotic matrix and spatial moment to capture the characteristics of Chinese words, the characteristic value is scaled to 0 to 1 interval output by formalization. With different font styles of 4808 Chinese words as the advantage point tree classifier database, and the advantage point tree classifier through the Euclidean distance range for Chinese word proximity search identification. Compared with the optical character recognition software of open source Tesseract-OCR, the identification results of 4808 Chinese words were compared. In the experiment, found that the establishment time of the advantage point tree classifier was less than 1 second, which was much less than the training of the deep learning model, and the average recognition rate of the database of the advantage point tree classifier with the new detail as the advantage point tree classifier reached 79% in the adjacent search for different type Chinese words, which was better than the recognition results of Tesseract-OCR Chinese words.
