| Due to the limited performance of monitoring equipment in public places,the quality of face images stored in surveillance videos is severely compressed,and most of them are low-resolution faces with few pixels,which makes it difficult to carry out some subsequent operations on low-resolution faces,such as face recognition,face alignment and other computer vision tasks.As a basic research in computer vision task,face super resolution reconstruction is to reconstruct low quality face image into high quality face image.The quality of the image reconstructed by the super resolution of face determines the performance of the subsequent relevant visual algorithms.In order to reconstruct high quality face images from low resolution face images,this paper proposes three methods of face super resolution reconstruction,and uses face recognition to verify the effectiveness of the proposed methods.The main research and contributions of this paper are as follows:(1)In this thesis,a face super resolution multi-scale feature fusion model based on deep learning is proposed.The model uses a multi-scale feature fusion structure to make full use of feature information at each level.Taking the maximum pooling layer as a jump connection can reduce redundant feature extraction and speed up network computing.The residual attention fusion module,which is the fusion of channel attention,spatial attention and pixel attention,can make the network task pay more attention to the image features which are helpful to the reconstruction of face image,such as face contour and facial features.The experimental data proved the validity of the super resolution multi-scale feature fusion model structure,and the ablation experiment also proved the advanced nature of the residual attention fusion module.The facial features reconstructed by the network of this module were clearer and more detailed.(2)In this thesis,a two-stage progressive difference complementary face super resolution reconstruction model is proposed.The model uses a two-stage processing approach to build a lightweight network that can extract more information from low-resolution face images.The first stage is the pre-processing stage,the input image is constructed into the high-low resolution image group which can be input into the second stage through the rough processing module.In the second stage,a difference complementary method is studied to mutually refine and supplement the features of the high-low resolution image,which can effectively reuse the features.In addition,a multi-kernel residual feature extraction module is developed to fully extract the global feature and local feature.The experimental results show that the asymptotic differential complementary network can perform better with fewer parameters.(3)In this thesis,a new face super resolution reconstruction model based on double generalized distillation is proposed.Most face super resolution models based on deep learning require a large amount of memory and neural units,which brings great challenges to deploying models on devices with limited computing performance.In order to solve this problem,this paper uses the method of double generalized distillation to obtain a small model which can be applied to mobile devices,namely the student network model in knowledge distillation method.Both the teacher network and the student network use generalized distillation method to obtain the privileged information of the original face image to assist the training,train the parameter weight of the student network by reducing the cosine distance of the middle feature image of the two networks,and supervise the training through pixel loss.The experimental data demonstrate the validity of the face super resolution reconstruction model based on double generalized distillation. |
PDF Full Text Request