Research On Tracking-based Face Detection And Face Recognition With Loss Function

Face recognition plays a critical role in surveillance and security systems.In this thesis,several novel methods are proposed for the face detection and feature extraction steps in the face recogntion system.Targeting at motion blur and occlusion between objects,we introduced a tracking-based deep detection network with attention mechanism.Based on the tracking results of the previous frames,a heat map is generated and fed into the deep detection network.Two-branch deep detection network is designed,whose two branches are merged with different heat maps.It makes network pay attention to different objects in two branches respectively.Meanwhile,in order to improve the face recognition accuracy,two loss functions are proposed.First,we develops a novel deep-based approach by introducing an adaptivelyweighted verification loss function.Our novel loss function can properly enlarge the margin between positive face pairs and negative face pairs from a global perspective.Thus a more reliable recognition model is obtained by minimizing the dissimilarity between same-person faces and maximizing the dissimilarity between different-person faces.Second,inter-class information is taken into further consideration.Based on the center loss function,we proposed the repulsive center loss.The distances between different class centers would be maximized by the new loss function.Each class center repels other class centers in each training iteration.The repulsive center loss can preserve the merit of center loss,while maximizing the inter-class variation.Besides,targeting at different kinds of variation problems,this thesis presents a novel network structure,called Extended Siamese Network.Synthetic face images under scale and eyeglass variations are generated for training process.Then the Siamese Network is adopted for the variation problem.The mapping layer and modified L2 loss are proposed in the Extended Siamese Network.On the one hand,the mapping layer can learn a mapping between two face images under variations.On the other hand,the modified L2 loss helps to maintain classification ability.