標題: 名片商標之擷取與辨識
Extraction and Recognition of Logos in Name Cards
作者: 張倍魁
Chang Bei-Kuei
李錫堅
Lee, His-Jian
資訊科學與工程研究所
關鍵字: 辨識;recognition;logo
公開日期: 2001
摘要: 在日常生活中名片被廣泛的用在保存個人資訊. 名片中的商標則是具有獨特的象徵意義.在這篇論文中, 我們設計了一個名片商標自動擷取與辨識的系統.因此我們可以辨別出名片之中的商標並把商標儲存到資料庫內. 我們系統的輸入是二值化過後沒有雜訊的影像. 系統包含兩個主要的模組:商標的擷取和商標的辨識. 再商標擷取的模組中, 我們首先使用了更快且更適合名片影像的改良式Run-Length Smoothing Algorithm (RLSA)已減少商標候選個數. 接著, 我們提出可能影響商標候選的五個調整: 對符合姓名欄位的商標候選減少商標可能性. 對符合文字行特性的商標候選減少商標可能性. 對含有較少連接單元之商標候選增加其商標可能性.對於含有大量連接單元的商標候選減少其商標可能性.對於面積小的商標候選減少其商標可能性. 最後我們選有最高商標可能性的商標候選為我們的商標.在商標辨識模組中, 我們設計了一個對名片商標具有高辨識率的辨識引擎. 我們用了三個個特徵值: contour directional features, crossing count features and peripheral background area features. 在我們的商標擷取實驗中, 我們選取了610張具有商標的名片. 成功率為82.9%. 在我們的商標辨識實驗中, 我們選取了187個公司共495個商標. 成功率為 96.9%
Name cards are used widely for keeping personal information in our daily life. Logos in name cards are designed to represent the corporation recognition mark. In this thesis, we design an automatic extraction and recognition tool for logos in name cards. Thus, we can identify logos in name cards and store logos in a database. The input of our system is binary name card image without noise. Our system contains two major modules: logo extraction and logo recognition. In the logo extraction module, we first use an improved Run-Length Smoothing Algorithm (RLSA), which is faster and more suitable for name card images to reduce candidates of logos. Next, we adjust the logo possibility as follows: Reduce the logo possibility of the smeared CCs if they satisfy the name field constraint, Reduce the logo possibility of the smeared CCs which are collinear in vertical or horizontal direction and have similar width or width in the orthogonal direction, Increase the logo possibility of a smeared CC if its size is similar to the size of the connected components in the original (un-smeared ) image, Reduce the logo possibility of smeared CCs which contain many small connected component in the original (un-smeared) image and Reduce the logo possibility of smeared CCs which are very small to calculate the score of the candidates. Finally, we choose the highest score of the candidates of the logo as a logo. In the logo recognition module, we design a good recognition engine which has a high recognition rate for logos in the name card. We use three features: contour directional features, crossing count features and peripheral background area features. In our logo extraction experiments, we select 610 images of name cards. The success rates are 82.9% in logo extraction. In our logo recognition experiments, we select 495 logo images of 187 companies. The success rates are 96.9% in logo recognition.
URI: http://140.113.39.130/cdrfb3/record/nctu/#NT900392061
http://hdl.handle.net/11536/68473
顯示於類別:畢業論文