| Video face recognition and deep learning are hot research topics in the field of pattern recognition, but they always encounter problems in practical applications. Deep learning has had some successful applications in terms of target detection and tracking, so that video face recognition research based on deep learning is carried out in order to automatically extract the sample characteristics which has a good power of expression.This thesis completes learning the convolution kernels, expanding the size of sample library and designing the deep network feature extractor by using relevant theories of deep learning to fully express the samples contained information. The thesis’s main work and innovation points are as follows:(1) This thesis proposes a convolution kernel learning algorithm based on denoise sparse autoencoder. We put sub-blocks of face images into it to get the parameters of the network, that is, all kinds of convolution kernels. The convolution kernels learned have stronger adaptability than traditional convolution kernels, it can extract the characteristics of the target better. We try to save time cost by making a study of choosing appropriate parameters including the convolution kernel size and training sample size.(2) This thesis puts forward the thinking of using convolution operation to expand data set. We use the convolution kernels learned to make convolution operation on face data sets. It effectively solves the network parameters fitting problem which dues to small training data set. Follow-up feature extractor uses pooling on convolution images to reduce the network complexity.(3) This thesis presents a design algorithm of the deep network feature extractor which has strong adaptability. It can learn out diagnostic characteristics according to the selected sample set. When it is combined with all kinds of classifiers, we can achieve good recognition result. It designs different layers to solve practical problems and complete face recognition. |
PDF Full Text Request