Face Image Processing And Recognition Methods Based On Subspace

In this paper, the main component of a face recognition system such as image preprocessing, face detection, feature extraction and face recognition is studied. In image preprocess portion, both the process for original image such as brightness and contrast preprocess, and that for extracted face portion from original image, such as unify process for size, zoom and lighting is discussed. In face detection section, a detect method based on artificial neural network and knowledge is discussed, and a fast detection arithmetic is put forward by learning the output characteristic of the face detector, and obtained a good performance. In feature extraction section, we generalized Yang's IMPCA. At last, face recognition method is discussed. In this section, a contribution matrix concept for face classification is first put forward. Its purpose is to determine the contribution matrix in terms of different contributions of each dimension of an eigenvector to the performance of the classification. And then a modified subspace classifier is designed by multiplying each original pattern vector with the matrix as a preprocessing method, and so an improvement of the recognition performance compared to the conventional subspace classifier on the Manchester face database is obtained. Secondly, to lower the mis-recognition induced from the overlapped subspaces and to increase the recognition precision, we employ the kernel method to implicitly fulfill a map from the pattern space to higher dimensional feature space so as to nullify the intersecting subspaces and to make the recognition problem easily-solved in the transformed space . Finally our experiments on the same face database verify the feasibility and effectiveness of the proposed method.