基於斷筆偵測之視訊手寫中文辨識系統

Video Handwritten Chinese Character Recognition System Using Stroke Segmentation

林倚鈴(Yi-Ling Lin)

指導教授 ： 謝禎冏

大同大學/電機資訊學院/資訊工程學系所/碩士(2010年)

摘要

近幾年來傳統輸入裝置漸漸被取代，利用手寫進行文字輸入更是日益普及。然而，手寫板及觸控板在使用上雖簡易，卻會造成攜帶上的不便。本論文提出一套以斷筆與線上模型為基礎之手寫中文輸入系統，其核心方法首先為利用指尖特徵抽取指尖位置，紀錄指尖連續移動軌跡，利用直線近似法原理將文字表示成一連續的直線段，依照筆劃轉折點作為特徵的判斷，由於深度資訊已經遺失，無法區分下筆和提筆的筆劃，因此以文字書寫習慣作為斷筆依據，並保存筆劃動態資訊。在斷筆規則中，我們歸納出幾種筆劃特性來區分提筆和下筆，主要分為五種，第一種為在書寫中文字的下筆上不會發生的方向筆劃，第二種為兩平行線之間的筆劃不會存在，第三種為不會出現與上一筆筆劃相反方向，第四種為左右部首之間的提筆，第五種為上下部首之間的提筆，利用以上規則可達到自動斷筆的效果。接著以方向、斷筆、前後筆劃所形成的夾角、及長度比例值三種屬性，組合成一連續的序列字串，在參考字串中，每個字我們建構了一組以上的線上模型，再以動態規劃之字串差異度比對演算法進行比對，達到線上即時文字辨識。在實驗中，我們建立了1000個中文字線上模型，邀請了5個人對每個字寫了3次，總共寫了15000個中文字，全部影像數為548726張，其中指尖追蹤準確率為98.88%，每秒可處理12.6張圖片，斷筆準確率為91.56%，文字辨識率為93.61%。證明本論文方法以指尖能有效輸入與以方向、斷筆、長度比與角度能正確比對識別文字。

關鍵字

斷筆 ； 指尖軌跡 ； 指尖追蹤 ； 手寫輸入 ； 手寫中文辨識

並列摘要

In recent years, traditional text input devices are changed gradually to handwriting. Though handwriting board and touchpad are easy to use, they are inconvenient to carry. This paper proposes a video handwritten Chinese character recognition system using stroke segmentation and on-line model. Firstly, the location of fingertip is extracted and the fingertip trajectory is recorded for recognition. The trajectory is straight line approximated by finding the turning points of strokes. Owing to the loss of depth information, it is unknown the user is going to write or move the fingertip. Therefore ,character writing habits are utilized to develop rules for pen-up and pen-down strokes classification. The main idea is to find impossible strokes which represent pen-up strokes for moving fingertip. The first case of pen-up stroke is from right to left or bottom to up. The second case is that the pen-up stroke would not exist between two parallel strokes. The third case is that the next pen-down stroke would not be of the opposite direction with the current one. The forth case is pen-up stroke between left and right character components. And the last case is pen-up stroke between upper on lower character component. These rules are used to segment the character into pen-down and pen-up strokes. In addition to stroke direction, stroke types, stroke length ratio, and angle between two consecutive strokes are also used to build the character online model as a four tuple continuous sequence of string. For characters written with multiple ways, we could build more than one online model for these characters. Minimum edit distance by dynamic programming is deployed to match the input character on-line string with stored online character models for recognition. In experiments, we build about 1000 Chinese character on-line models. The recognition system is tested with five persons writing each of the 1000 Chinese characters three times. There are totally 548726 images with camera of capturing speed 30 frames per second. The accuracy of fingertip tracking is 98.88% with processing speed 12.6 times per second. The accuracy of pen-up and pen-down stroke segmentation is 91.56% and the accuracy of character recognition is 93.61%. These results demonstrate that the proposed method could be used to input characters by fingertip efficiently.

並列關鍵字

handwritten Chinese character recognition ； stroke segmentation ； fingertip trajectory ； fingertip tracking ； handwritten character

參考文獻

[2] 林柏宏，“視覺式計人次暨性別辨識系統-以數位廣告看板為例”， 私立大同大學資訊工程研究所碩士論文，2009年8月

[8] K. Morimoto, C. Miyajima, K. Itou, and K. Takeda “A virtual button interface using fingertip movements,” Machine Learning and Cybernetics, vol. 4, pp. 2089-2093, August 2007

[9] L. Rabiner, “A tutorial on hidden Markov Models and selected applications in Speech Recognition,” IEEE Trans on ASSP, vol. 77, no. 2, pp. 257-286, February 1989

[11] J. A. Montero and L. E. Sucar, “Feature Selection for Visual Gesture Recognition Using Hidden Markov Models,” Fifth Mexican International Conference, pp. 196-203, September 2004

[12] L. Nianjun, B. C. Lovell, P. J. Kootsookos ,and R. I. A. Davis, “Model structure selection and training algorithms for a HMM gesture recognition system,” Frontiers in Handwriting Recognition, pp. 100-105, October 2004