Research On Deep Learning Face Recognition Based On Second-order Pooling And Over-complete Representation

Face recognition has been one of the research hotspots in both academia and industry.In the past few decades,researchers have proposed a series of face recognition methods.However,these methods cannot achieve good recognition results due to the external disturbances,such as pose,facial expression,illumination and so on.Convolutional neural networks(CNNs)have enough capacity to represent the complex variations of samples by learning features hierarchically.Therefore,CNNs have made outstanding achievements in face recognition.The common models mainly adopt the average/max pooling layers in designing network architectures.However,these layers only capture the first-order statistics of input features,which limit the learning capacity of models.In addition,faces contain not only the global discriminant information,but also contain the local discriminant information.So the performance of single model is limited to the loss of discriminant information.To overcome these issues,this thesis proposes second-order pooling based deep learning for face recognition and over-complete representation based deep learning for face recognition.The core component of the first method is Second-Order Pooling Convolutional Neural Network(SOPCNN),which can improve the capacity of traditional models.Firstly,the face image is transformed to a feature map by several convolutional layers and pooling layers.Then the feature map is decomposed into a set of local features,and the corresponding result matrix is obtained by the outer product operation for each local feature.Finally,the results are used as the input of the average/max pooling layer.Pooling layers can learn more robust representation by extracting second-order statistics of the input features.The core component of the second method is Multi-Region Convolutional Neural Network(MRCNN),which can efficiently learn more complete and more robust representation.Firstly,a full face image and a set of face regions are used as inputs.The shared layer is used to transform the face image into a feature map,which is composed of several convolutional layers and a pool layer.Then,the mapped region is extracted from the feature map for each face region.A separate network will learn features of the mapped region.Finally,the over-complete representation of a given face is obtained by extracting and combining features of all regions.Compared to some available methods,the MRCNN model has better computational efficiency by sharing hidden layers.The proposed algorithms are evaluated on CASIA-Web Face and LFW.The experimental results demonstrated that the proposed algorithms can improve the effectiveness and robustness of face representation.