在人臉辨識的研究中，大多數人是以特徵擷取的主成份分析法 (principal component analysis, PCA) 及線性鑑別分析法 (linear discriminant analysis, LDA) 來辨識人臉。這二種方法在多視角 (multi-views) 辨識上各有其優缺點，本論文即整合這二者的優點，提出一個多視角分類法 (先以線性鑑別分析法做臉部視角分類，再以主成份分析法就選定的類別做辨識)，以提高不同視角臉部辨識的辨識率。為了執行效率，我們也利用離散小波轉換 (discrete wavelet transform) 做多重解析度影像分解，一方面降低影像解析度做分析，另一方面以高頻資訊擷取臉部明顯區域做分類。 在實驗中，我們以40人所組成的400張多視角人臉影像來測試本系統的效能；其中200張影像做訓練，另外200張做測試。我們共做了三種實驗：(i) 多視角分類法，(ii) 直接以主成份分析法做辨識，及 (iii) 直接以線性鑑別分析法做辨識；並比較這三者方法在多視角人臉辨識上的效能。實驗結果顯示在各種不同樣本集合測試下，多視角分類法的辨識率都優於主成份分析法；另外在訓練樣本數少於30人的情形下，多視角分類法也比線性鑑別分析法好。 In the study of face recognition, PCA (principal component analysis) and LDA (linear discriminant analysis) are commonly used for recognizing human faces. Both methods have individually advantage and disadvantage for multi-view face recognition. We here proposed a recognition system to improve the recognition rate for multi-view face recognition by integrating the advantages of PCA and LDA. In the proposed recognition system, the LDA is used to classify facial views, and then PCA is used to recognize face images in a specified class. For efficiency, the discrete wavelet transform is used to reduce the image resolution for recognition as well as to extract the face region for classification based on the data in the high-frequency subbands. In our experiments, 400 multi-view images of 40 persons were captured for training and test. We compared the performance of the three methods: (i) multi-view recognition, (ii) PCA, and (iii) LDA for multi-view face recognition. With the same face samples, the performances of multi-view recognition method are always better than that of PCA and LDA (at most 30 persons).
