| 摘要: | 在數位化時代，不論是入出境管理、網路轉帳、居家或公司大樓門禁等，都需要進行身分識別，以確保個人權益、資訊安全。傳統的身分認證方式，已經不敷使用，例：身分證件、印章認證……。當前生物辨識技術，是最具方便性與安全性的認證方式，例：指紋認證、人臉認證、視網膜認證……。
在生物辨識技術中，“人臉認證” (Face Authentication / Face Verification) 不需要近距離接觸、不需要使用者主動參與或被動配合、生物特徵採集過程不會產生任何不適感，對於一般大眾具有較高的接受度，應用層面也較為廣泛。本論文提出使用演算法LGP (Local Gradient Patterns) [1]、Hybridization (LBP and LGP) [1] 採集人臉的生物特徵，搭配AdaBoost (自適應增強)技術訓練出人臉認證的參考模型。
LGP演算法分別計算鄰近像素的梯度值，並使用鄰近像素的平均值做為中間像素的門檻值，如果鄰近像素的梯度值小於門檻值，則該鄰近像素標示為0，反之，則為1。相較於LBP演算法，LGP能夠維持物體邊緣具有較為健全的亮度、色彩飽和度。利用最佳的局部轉換特徵具有最小分類誤差性質，Hybridization透過AdaBoost把LBP、LGP的特徵進行融合。融合後的Hybridization，同時擁有LBP、LGP的優點，能夠進一步提升效能。
本論文於實驗階段，分別採用不同數量的人臉正、負樣本，做為訓練參考模型的來源影像，以確認樣本數量對分類器的影響性。同時，也使用CMU PIE Database的68個人進行身分認證，確認每個人的認證效能差異。實驗結果顯示本論文所提出的LGP、Hybridization演算法及驗正方式，運用在身分認證系統中，能夠有效提升準確率，並降低身分認證被仿冒風險。
;In the digital era, we have to identify everyone to protect individual rights and information for bureau of entry and exit, network transactions, entry access control at home or corporate, etc. Traditional identity authentication methods have been insufficient for our needs, for example ID card, seal certification, etc. Currently, biometric technology is the most convenient and secure way on fingerprint authentication, face authentication, retina authentication, etc.
In biometrics, face authentication not only has the highest acceptance of general public, but also the wide range of applications. It has three advantages 1) no need to touch 2) does not require user involvement or cooperation 3) the biometric collection process will not have any discomfort. Therefore, we propose the novel local transform feature: local gradient patterns (LGP) [1] and hybridization feature [1] that combines LBP, LGP by means of the AdaBoost method in face authentication (face verification). It will transform the face images into the LGP, and hybridization feature images. Then face authentication model was trained base on feature images and AdaBoost algorithm.
LGP will calculate the neighboring gradient of a given pixel and its average of neighboring gradient. Then the average of neighboring gradient was set to center pixel. If neighboring gradient is greater than center pixel, LGP assigns one and zero otherwise which makes the local intensity variations along the edge components robust. According to the best local transform feature having the lowest classification error, LBP and LGP feature are fused by AdaBoost for hybridization of local transform features. This hybridization makes face detection performance robust to changes in global illumination by LBP, local intensity changes by LGP.
In the actual experiments, we utilize a different number of positive samples and negative samples training face authentication model and the accuracy under various sample numbers are demonstrated. In addition, CMU PIE database is applied in our experiments. Experimental results show that our LGP and hybridization could improve accuracy and reduce the risk of counterfeit identity in terms of face authentication. |
