Application Of Super-resolution Reconstruction Based On Deep Learning In Face Recognition

With the development of deep learning,many technologies in the field of computer vision have made rapid progress,and the applications based on deep learning models greatly facilitate our lives.Among them,The face recognition technology based on biometric has been widely used.Residential communities,traffic routes,security departments and other scenarios are equipped with cameras,forming a more complete video surveillance system,image data acquisition becomes easier,providing a sufficient source of information for intelligent analysis technology.However,the image equipment of most scenes have low quality,and is affected by its own hardware capabilities,illumination,environmental noise,and target distance during data acquisition,resulting in a greater degree of damage to the final image data,which is manifested by blurred images and the missing details of face,leading to the poor recognition accuracy.In order to solve the problem in this case,this thesis will research the superresolution reconstruction technology based on deep learning and combining it with the traditional face recognition process to restore the feature information in low-resolution face images as much as possible and improve the recognition accuracy.The main research contents are:Firstly,exploring and comparing the performance of different activation functions in the deep learning network,and proves that the pre-activated residual unit has a better feature mapping mechanism in the deep network than the original residual unit.Secondly,analyze the existing deep learning network,combine the advantages of VDSR and ResNet models,propose a structure of the residual unit,and introduce it into the super-resolution reconstruction network.The layer features can flow to any deep layer and give the network better backpropagation characteristics,preventing the gradient vanishing problem during training.Using ELU instead of ReLU as the activation function,the network node can make a certain response to the negative input during training,the update direction of the weight becomes bidirectional,suppresses the oscillation generated before convergence,and can avoid the problem of “dying ReLU”.Thirdly,combining the existing face recognition algorithm with the superresolution reconstruction algorithm proposed in this paper to form a new framework,and exploring the impact of super-resolution reconstruction algorithm on facial feature extraction by feature dimension reduction and visualization experiments.Finally,The algorithm framework of this paper is tested by LFW dataset and a large number of image containing the low-resolution face from actual scenes.The experimental results show that the proposed super-resolution algorithm can improve the accuracy of the existing face recognition algorithm by 9.8% in low resolution scenarios.