| Human face recognition is a hot research topic in image processing, pattern recognition and computer vision fields. It has tremendous potential applications in criminal hunting, law enforcement, employee attendance checking and video surveillance. For the last two decades, many effective face recognition algorithms has been developed, while lighting variation is still one of the main difficulties hindering this technology to go from laboratory to practical use. In this thesis, the problem of face recognition with good illumination robustness is deeply investigated, and three approaches as follows are put forward.Firstly, a method based on the traditional Principal Component Analysis (PCA) algorithm, but with face images being illumination compensated at preprocessing stage, was proposed. In order to remove the influence of lighting changes, the low frequency components of a face image are discarded using wavelet multi-resolution analysis method, and only the high frequency components are used for the reconstruction of a face image. Experimental results show that the face recognition rates of PCA method on the reconstructed face images can be improved a lot.Secondly, a method based on Binary Edge Map (BEM) and Support Vector Machine (SVM) was presented. The BEM of a face image was extracted with Locally Adaptive Threshold (LAT) algorithm. Experimental results on AR face database and Yale face database show that this proposed method can achieve better illumination robustness than the Doubly Modified Hausdorff Distance method (M2HD) and PCA method.Finally, a method based on the fusion of binary edge feature and grayscale feature was proposed. Edge feature can result in good lighting robustness, but may lead to the deterioration of face recognition methods in the cases with facial expression changes. Experimental results show that the method based on the proposed feature fusion scheme can improve both the illumination and expression robustness, making it more applicable. |
PDF Full Text Request