文字擷取及利用筆劃和視覺文字辨識

這篇論文會介紹從一般普通的招牌上取出所需要的文字及辨識。一開始會先從普通的圖片讓他做灰階轉換，對比強化等動作。使得圖片會只剩下黑色與白色。當剩下黑色與白色之後，我們會去尋找黑色的連通圖案。並且把它視作是一個字元。當取得我們所需要的字元後接下來會換到辨識的部分。辨識的部分我們分為兩個階段。第一個階段會使用筆劃分解的方式來解讀該字元是什麼。進行筆劃分解的時候，因為細化演算法有時候會使字元產生一些變型所以當分解完筆劃後需要做線斷合併的動作。當作完線斷合併後，筆劃分解主要把它分解成以交叉節點和終端節點來判斷該字元是什麼。不過如果在第一階段分解階段的時候，得到預想之外的結果的話則會使用第二階段VisualWord來做辨識。其辨識的方法是先把字元圖案A到Z經過特徵萃取及分類的階段之後，建構出辨識字元用的資料庫。之後當需要使用這個方法來判斷字元的時候，同樣的會對該字元做特徵萃取及分類。然後比較該字元與資料庫裡哪一把資料相似度最高就把他認定為該字元。

關鍵字

文字辨識

並列摘要

Optical Character Recognition (OCR) [1] is the most famous method to solve the text recognition. It has extremely high accuracy in text recognition, but it needs a large amount of computation and database. Therefore, we propose a method that uses smaller computation and database. Furthermore, we also achieve a high accuracy in recognition. This paper shows how to extract text from an image and use the visual word to recognize the text. First, we describe an approach to translate an image in sign board to a gray-scale image. After gray-scale translation, contrast enhancement, etc., we can get the binarized image, search the connected component, and find the character in this binarized image. Second, we recognize the character. We use stroke analysis and visual word. In stroke analysis, we decompose the character and combine lines. We can get endpoint number ,crosspoint number,line number..etc. We use these informations to find the corresponding character. If we can not find the corresponding character, we use visual word to recognize the character. In visual word,we build a database to recognize the text in the image. We use Harris corner detector [2] to find the interesting region in the binarized image. Then we use SIFT [3] to extract the feature from this region. After using SIFT to get the feature, we will have 128-dimensional vector. Next, we use K-means clustering to label every region in the data image. Therefore, for each character, we can collect a character-vector using its region IDs. After building the database, we use two steps to recognize the character. First is stoke analysis. If the character is not sure, we follow the step 2. Second step,calculate the similarity between the character of the input image and each character in the database using their character-vector. Hence, we can recognize the text in the input image. iv

並列關鍵字

word recognizition

參考文獻

[1] L. Eikvil. Optical character recognition. December 1993.

[2] K.G. Derpanis. The harris corner detector. 2004.

[3] D.G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal

[5] L.R. Rabiner. A tutorial on hidden markov models and selected applications in speech recognition.

In Proceedings of the IEEE, pages 257–286, 1989.

國際替代計量

文字擷取及利用筆劃和視覺文字辨識

主題瀏覽