Research On Key Algorithms Of Face Detection And Face Recognition In Video Surveillance

The role of video surveillance in the security field is increasingly significant.At present,the understanding of video content basically relies on manual judgment,so how to effectively utilize surveillance video has become a research hotspot.Although the face detection and face recognition in intelligent video surveillance systems is developing in full swing,the performance of face detection and face recognition in video surveillance cannot meet the expected requirements due to the influence of illumination,face pose,multiscale and so on.Deep learning is widely concerned due to its excellent feature expression ability.In this thesis,based on the deep learning,the specific problems encountered in face detection and face recognition in video surveillance are studied and improved algorithms are proposed.The contributions of the thesis can be summarized as follows.1.Based on the representative algorithm about multi-scale face detection S~3FD and the idea of object detection algorithm FPN,the detection ability on small-scale face is enhanced from the perspective of fusion of different scale features.The data augment sampling and Max-out classification strategy further improve the ability of the face detection model to detect faces at different scales.2.In view of the slow detection speed of S~3FD,the network structure has been improved and compressed to improve detection speed.The application of the C.ReLU activation function allows the shallow convolutional layers of network to complete all operations with half of the convolution kernel.In order to further increase the speed of the face detection,the amount of computation is further reduced by compressing the number of convolution kernels in network.3.In response to the lack of video surveillance datasets for face recognition,a number of Asian ethnic data and data for video surveillance were collected through collaboration with related projects and the network.And the dataset is augmented by the multi-pose data synthesis based on 3D model and the improved face attribute translation method based on Generative Adversarial Networks(GAN).4.Aiming at the problem that the performance of the state-of-art face recognition model on the public test set is seriously degraded in the video surveillance,face recognition algorithms and transfer learning algorithms are studied to improve a Siamese Network algorithm based on metric learning for face recognition model transfer in the video surveillance.