Research On Face Recognition Based On ICA

As the development of the society, there are increasing demands in automatic identity certification . Since some biological characteristics are intrinsic, stable and strongly different from each others, face recognition is becoming attractive in pattern recognition and image processing.Generally, face recognition consists of three parts which are preprocessing, feature extraction and classification. Aim to the changes of expression and illumination of face, we mostly study on how to extract more useful face information.The preprocessing work includes histogram equalization and whiting, which effectively improve the image quality, decrease the computation complex and speed up the algorithm's convergence.In the part of feature extraction, firstly ,the paper uses PCA (Principal Components Analysis) to reduce the dimension of train images, which eliminate the second gradation redundancy information of them, and make PCA subspace as the input data of ICA (Independent Component Analysis)algorithm. It shows that the training time will be increasing exponentially when the train samples are large. This process can decreases the computation effectively and improves recognition rate.Secondly, considering the specificity of face images, we adopt the FastICA (Fast Fixed-point ICA) algorithm based on negentropy to extract the face feature information. The method has a excellent partial feature expression ability. It has good astringency and less computation as well, by using fixed point iterative adaptive study method. In addition, the paper proposes the kernel independent component analysis method,which is based on the nonlinear function space, implementing the nonlinear transform with kernel functions instead of inner-product between two vectors. It has been proved that the method has more flexible and good robustness for the change of the express and illumination .Finally, as ICA face feature space is concerned, we use the most direct method-- Nearest Neighbor classifier, which is widely used, to classify recognition in the paper. The algorithms involved in the paper have been simulated by Matlab using Yale face data and AR face data, and have showed good recognition rate.