| 摘要: | 隨著資訊安全逐漸受到重視，利用生物特徵作為身份認證技術的安全系統扮演了重要的角色。在過去的幾十年裡，已有許多生物特徵被廣泛的應用在個人認證系統，例如臉、指紋，以及虹膜等。而在過去幾年也有許多文章討論將單一的掌紋或者掌背血管當作特徵，但這些文章所描述的方法在使用上均受到某些限制，例如，擷取手掌影像時需要一個固定手掌的裝置，確保手掌不會因為位置及方向而萃取出不同的手掌區域；需要足夠的光線來擷取影像。 本論文中提出一個新的方法來避免上述的種種限制，除了可以利用非固定式的裝置，並可以同時擷取掌紋影像以及掌背血管的紅外線影像。此外，從這些蒐集的影像資料中擷取出感興趣的區塊，並利用此論文中提出之新的特徵擷取方法來獲得影像中的特徵資訊，最後我們結合這兩種生物特徵來增加此認證系統的辨識率。 我們分別利用兩個生物認證技術方法來區分是否為本人，分別是樣版比對以及支援向量機兩種。首先，樣版比對計算線性相關係數，再利用參考樣版來比對，以作為兩張不同的掌紋影像間相似度比較的依據。第二種方法採用了支援向量機來辨識，我們使用輻狀基底函數當作核心來訓練影像資料，最終利用交叉驗證得到了98.2%的準確率。實驗結果驗證了我們方法的可靠性、可行性及可適性，我們相信此論文中提出的論點是個值得注意的方法。 Due to the explosive growth and demand of information security, biometric features based personal identification systems gradually dominate identification techniques in many applications. There are many physiological features, such as faces, fingerprints, and iris images, which have been extensively studied for personal verification purpose in the past few decades. In the past, many literatures discussed biometric verification by only using either palmprint features or palm-dorsa vein-patterns features. Both of them are constrained by some certain limitations (e.g., utilization of fixed pegs to constrain the palm position while acquiring palm images, and requirements of adequate lighting conditions) which hinder the practicality of applications. In this thesis, a novel method is proposed to remove the limitations imposed by docking devices so that visible thermal images can be captured from the pegs-free platform. Furthermore, the palm features are successfully extracted from interest of interest (ROI) by using the novel proposed histogram of iterative thresholding (HIT) technique. Finally, we combine the two palm features, which are principal palmprints and palm-dorsa vein-patterns to enhance the accuracy rate. In our work, template matching and support vector machine (SVM) are designed to verify the query images. First of all, linear correlation function is adopted in template matching, and then query image is matched with the reference templates to measure the similarity between the two different palm images. Secondly, we train the data with the selected RBF kernel in SVM. The k-fold cross validation is used to obtain a 98.2% accuracy rate. Experimental results demonstrate that our proposed algorithm is reliable, feasible, and adaptable in practical applications. |
