應用字型生成資料開發環境中文字辨識系統

Abstract

近年來，隨著人工智慧的快速發展，深度學習（Deep Learning）的技術也隨之蓬勃發展，並廣泛應用在各個領域，包括中文字元影像辨識（Chinese Character Recognition）。 本研究的目的在改善中文漢字之辨識模型建立問題，利用現有電腦系統內建的字型資源來產生文字影像，再經由一系列的影像處理來模擬真實環境中的影像型態，並調整影像內文字本體部分數值，使得在使用機器學習中的卷積神經網路（Convolutional Neural Networks）之技術時能更有效學習到文字架構特徵而非邊界像素點分布之特徵。 經由實驗結果顯示，使用本研究方法在現代報紙與民初晶報等印刷文件之辨識準確率分別為97.66%與78.21%，在CASIA 公開中文手寫測試集內達到63.15%之辨識準確率，以及在針對ICDAR-2019年ReCTS (Robust Reading Challenge on Reading Chinese Text on Signboard)競賽內之測試資料集，在使用官方提供之訓練資料額外加入本研究方法所產生之文字影像一同訓練，達到91.26%的辨識準確率，上述所提及之辨識表現優於現有OCR系統及方法。

In recent years, with the rapid development of artificial intelligence, deep learning technology has also been widely applied to various fields, including Chinese Character Recognition. The main purpose of this paper is to solve the problem of Chinese character recognition model building. By using the existing Chinese font resources in computer system to generate text images, and then use a series of image processing to simulate the image in the real environment and adjust the pixel value of text in image. That makes it more effective to learn the features of the text structure rather than the characteristics of the boundary pixel distribution when using the technology of Convolutional Neural Networks in machine learning We conduct our experiments with newspaper and the Jing Newspaper, the CASIA handwritten Chinese character public test set and the Chinese character of ICDAR-2019 ReCTS race testing dataset. The results show that the model and method we proposed in this paper can reach the accuracy of 97.66% on newspaper, 78.21% on Jing Newspaper, 63.15% on handwritten, and 91.26% on ICDAR ReCTS. Compared with the existing common OCR recognition software, our method can improve the accuracy.