由於傳統區別分析中常用 PCA 進行前處理，但 PCA 前處理可 能破壞原始資料的結構，使區別能力下降，為了降低 PCA 的負面影 響，本篇論文提出將原空間的資料透過不同的 kernel Methods 映射至 kernel 空間，目的是使資料在特徵空間中更具區別性，再於特徵空間 使用 PCA 將主成分提出；並配合最近特徵線轉換法 (NFLE) 進行區 別分析，以提昇辨識率。 在實驗部分，我們使用 CMU 資料庫做人臉辨識，針對不同的 樣本數與不同的維度下進行效能驗證 NFLE , Linear Kernel+NFLE , Guassian Kernel+NFLE , Polynomial Kernel +NFLE 四種不同的演算 法； 由實驗數據可得，雖然 Kernel Methods 在低樣本數的辨識率低 於 NFLE，但隨著樣本數增多，Kernel Methods 的辨識率則會高於 NFLE。; In traditional discriminant analysis, PCA is usually applied for data preprocessing. However, PCA may bring damage to the topology of original data and hence decrease the discriminability. To remedy this problem, the Kernel methods are adopted to transform the data set from original space to feature space for enhancing the discriminability in this study. In the kernel space, the PCA is then applied to extract the principal component data and remove the noises. After the PCA process, the NFLE algorithm is applied for discriminant analysis. In the experiments, Linear Kernel+NFLE , Guassian Kernel+NFLE, Polynomial Kernel+NFLE algorithms are implemented for face recognition. In our work,the CMU face database is used for evaluating the performance of the proposed methods in different training samples and dimensions. Experimental results reveal that the recognition rate of the proposed kernel based method is lower than NFLE under few training samples. When the number of training samples increases, the proposed kernel based method outperforms NFLE.
