Development of PLS-based EMD for face recognition
針對此種問題，過去有許多演算法被提出，其中以EMD為基礎的技術在近年來漸漸受到重視。研究學者使用不同的內插法來實作BEMD, 並應用所提出的方法來對臉部影像作前處理以達到可靠的辨識結果。 本論文嘗試以Tri-PLS對臉部影像資料庫建樣板，並以此樣板來取代典型BEMD方法的內插過程，再進一步檢視此種方法的有效性。
結果顯示所提出的方法PLS-EMD在辨識率上優於此兩種方法。尤其在使用PIE database作評估時，Eigenface的LOOCV辨識率約只有20%，而PLS-EMD的LOOCV辨識率可達到九成，如此大的差距說明了PLS-EMD相較於一般的辨識方法，對影像間存在不同光線條件變化具有可靠的辨識能力。另外再進一步與典型的BEMD方法作比較，在實驗過程裡，BEMD及PLS-EMD的threshold值分別為0.01及0.1，PLS-EMD所採用的樣板個數為30個。結果同樣顯示所提出的方法無論是以repeated random sub-sampling validation 或 LOOCV做評估，在辨識率上均優於典型的BEMD作法。而在計算效能上，若以PIE database作評估，BEMD在特定停止條件下，平均每張影像計算時間約需50秒; 而PLS-EMD在此條件下僅需花費約7秒。由此可知其計算效能的差異。|
Face recognition is an important application in the field of pattern recognition. Researchers devote themselves to enhance the reliability of the face recognition system in the past thirty years and many of the related algorithms emerge as the times require. As one of the major challenges in face recognition mentioned in literatures, some recognition algorithms will probably misclassify faces when different illumination conditions are present among face images. As one kind of feature extraction methods, feature-based methods extract features to distinguish different faces by solving eigenvalue problem; however this kind of method such as Eigenface and Tri-PLS suffers from the problem mentioned above. The feature-based methods take the whole regions of face image into consideration, and therefore the recognition rate will degrade significantly when the different illumination conditions are present in face images. Some algorithms have been proposed to deal with this problem in the past. As one kind of solutions, EMD-based methods have received significantly attention in recent years. Researchers adopted different interpolation methods to implement BEMD, and then applied the proposed methods to preprocess face images in order to enhance the recognition accuracy. In this study, a different approach which replaces the interpolation process of BEMD by selecting templates is proposed. The templates of the face database are obtained by Tri-PLS. Yale and PIE databases are applied to evaluate the recognition rate of the proposed method. The compared methods include Eigenface and Tri-PLS. Simulation results show that the recognition rate of the proposed method (i.e. PLS-EMD) is better than that of the two methods. The recognition rate of Eigenface and Tri-PLS evaluated by leave-one-out cross-validation are only about 20 and 40 percent respectively, and that of PLS-EMD is about 90 percent. The great disparity between the two recognition rates indicates that PLS-EMD has a reliable recognition ability to resist different illumination variations between images compared to other general methods. Another comparison between BEMD and PLS-EMD is also given. The recognition results evaluated by sub-sampling validation and Leave-one-out cross-validation also shown that PLS-EMD has better recognition rate than that of BEMD. As for computing performance, the average computation time per image of PIE database computed by BEMD which adopts specific stop criteria is about 50 seconds; however, it only takes about 7 seconds for PLS-EMD under the same condition. It is obvious that the computing performance of PLS-EMD is better than that of BEMD.
|Appears in Collections:||Thesis|
Files in This Item: