本論文提出特徵線度量學習框架，應用於身份辨識，包含有人臉驗證(face verification)、人臉辨識(face recognition)與行人識別(person re-identification)。傳統進行圖形識別會使用度量學習進行樣本之間的相似度的計算。進行影像識別時，會使用高維度特徵向量表示各影像的特性，也相對需要大量運算資源。為了解決高維度產生的高運算量問題，鑑別分析方法(discriminant analysis)被應用於降低特徵維度並提高資料的鑑別度。常見的鑑別學習方法有線性鑑別分析及區域保留投影等方式。有別於傳統使用樣本點到樣本族群中心或樣本點到樣本點計算特徵向量各個維度的權重值，本論文提出以樣本點到樣本特徵線(feature line embedded)取代點到中心或點到點的計算方式。本論文第一個主題為人臉驗證，提出特徵線嵌入偏頗鑑別方法(FLE-BDA)，計算測試影像與指定影像的相似度，驗證是否為相同身份。第二個主題為人臉辨識，本論文結合核空間(kernelization)與最近特徵線提出核空間最近特徵線嵌入分析(KNFLE)，測試影像與資料庫影像經過投影降維後，以歐氏距離比較兩者的相似度，藉以辨識測試影像的身份。第三個主題為行人再識別，則使用特徵線方式計算樣本之類的權重關係，同時計算樣本的機率分布，嵌入馬氏距離。測試樣本與資料庫樣本相似度，用來計算不同攝影機之間的行人影像相似度。上述問題人臉驗證及辨識使用Yale B、ORL及CMU PIE人臉資料庫進行效果驗證。行人再識別問題我們使用VIPeR與QMUL GRID行人影像資料庫與相關研究進行比較。;This dissertation proposed a metric learning framework based on feature line embedded for person identification. In order to avoid the high computational complexity required for high dimensional features, the discriminant analysis methods was used to transform data from high dimension space to low dimension space. The discriminant analysis can reduce the feature dimension while increasing the class separation ability. The point to line (P2L) strategy can be used to find the effective and discriminant transformation in eigenspace. So that, P2L was integrated into biased discriminant analysis algorithm to solve face verification firstly. Secondly, we used P2L in kernel space to deal the face recognition problem. Thirdly, the P2L based quadratic discriminant analysis was used to solve the person re-identification issue. In face verification, the biased discriminant analysis focus on positive samples. The projection matrix decrease the distance between positive samples and positive center. Meanwhile, the distance between negative samples and positive center were increased. We used the concept of biased discriminant analysis to propose the P2L based BDA to solve the face verification issue. In face recognition, since face recognition is a multi-class problem instead of two-class in face verification. Therefore, we apply the feature line embedded method preserving the local structure information in kernel space for increasing the recognition rate. In person re-identification, the quadratic discriminant analysis based on P2L consider the data distribution and the local structure information at the same time, the performance of person re-identification could be raised. In the experiments, we discuss the effects of various discriminant learning method on face verification, face recognition, and person re-identification. In face verification and recognition issue, the toy samples experiments showed the data distribution of discriminant analysis method. The real facial image database: Yale B, ORL, and CMU PIE face database used to evaluate the performance of face verification and recognition. The equal error rate of face verification was used to evaluate dimensionality reduction method. The recognition rate showed the effect of face recognition. The VIPeR and QMUL GRID pedestrian dataset were used to evaluate the efficiency of person re-identification algorithm. The experimental results showed our proposed P2L strategy based quadratic discriminant analysis was the best.
