於名片辨識系統中以連接元件為基礎的版面分析技術

Abstract

在這篇論文中，我們利用連接元件的幾何特性為基礎，設計了一個在名片辨識系統上，做為版面分析與產生文字行的系統。對於產生文字行方面，我們針對於塗抹所容易造成的缺點去分析。以連接元件為基礎，用Split-and-Merge的方式去產生文字行。並利用連接元件與文字行的幾何特性去提升文字行產生的正確率。用以降低名片辨識系統因為文字行產生錯誤所造成的辨識與後處理的錯誤。 在版面分析部分，我們利用連接元件的幾何特性為基礎來分析名片。我們利用全形與半形的連接元件大小以及各種語言的書寫方式定義出特徵向量。接著我們利用特徵向量的幾何特性，如向量的長度和方向性，來將名片版面歸類到橫式橫寫、橫式直寫、橫直混合、直式直寫、直式橫寫、直式混合六大類中。並依據分類結果提高文字行產生的正確率。 在文字行產生方面，我們測試了六大類共720張名片。文字行分析正確的有98%。而在版面分析方面，我們利用本論文所提方式，於484張名片中，將82.4%名片的版面正確分析出來。

In this thesis, we design a layout analysis and textline generation system on business card recognition based on geometric properties of connected components. In the textline generation, we analyze the disadvantages of smearing and propose a modified method base on connected components. We use the recursive splitting and merging to generate textlines. Then we use the geometric properties of connected components to increase the accuracy of textline generation, reduce the time of post-processing, and decrease the mistakes which many made by erroneous textline generation from the business card recognition system. In layout classification, we analyze the business cards based on the geometric properties of connected components. We use the full size and half size of connected components and the writing style of languages to define the characteristic vectors. Then we classify the layout of the business cards based on the geometric properties of characteristic vectors, such like the length and direction of characteristic vector. We classify the business cards into six styles: horizontal writing of horizontal style, vertical writing of horizontal style, mixed writing of horizontal style, horizontal writing of vertical style, vertical writing of vertical style, mixed writing of vertical style, and use the information of style to increase the accuracy of textline generation. In textline generation, we test 720 business cards in six styles. The correctness of textline generation is 98%. In layout classification, we test 484 business cards in six styles, and the average accuracy of layout classification is 82.4%.