Research Of The Low Resolution Face Recognition

In recent years,face recognition algorithms based on deep learning have been increasingly used in security,payment,access control and other fields.In the high-resolution face recognition task,the deep learning model has achieved a very high recognition rate and has a high algorithm stability.However,in the scenes of building security,criminal arrest,lost child recognition,etc.Face images tend to have lower resolutions,which leads to a reduction in algorithm recognition rate.Regarding the problem of paired recognition of such low-resolution faces and high-resolution faces,a large number of super-resolution and low-resolution face recognition models have emerged in recent years,but such models still have low super-resolution accuracy and the recognition rate is not high enough.In response to these problems,based on the generated adversarial network,this paper has carried out research on face super-resolution and low-resolution face recognition algorithms.The main research results are as follows:(1)A super-resolution algorithm based on generative adversarial networks is proposed.The proposed algorithm includes coarse super-resolution network,fine encoder,a priori feature estimator,fine decoder and discriminator.In the discriminator,this paper introduces Patch GAN structure and multi-scale discriminator structure.In the design of the loss function,this paper introduces a multi-layer adversarial loss function,and also introduces the VGG-16 model to extract identity features,and introduces a multi-layer perception loss.For pixel loss,this paper uses MAE loss to replace the traditional MSE loss.Through qualitative and quantitative experiments,the effectiveness and feasibility of the proposed algorithm in face super-resolution tasks are finally verified.(2)For non-surveillance scenes,a low-resolution face recognition algorithm based on super-resolution is proposed.The proposed algorithm is divided into super-resolution network and face recognition network.In this paper,Arcface face recognition model is selected as the pre-trained model.A data filtering algorithm based on graph cut is proposed,which selects high-quality pictures from the VGGFace2 dataset and performs super-resolution,expands the obtained super-resolution images to the VGGFace2 dataset to obtain the training set,and fine-tunes the pre-trained model.It enables the model to map super-resolution images and high-resolution images to a common feature space with identity consistency.Experiments show that the proposed algorithm canachieve a high recognition rate in low-resolution face recognition tasks in non-surveillance scenes.(3)An improved super-resolution network is proposed for surveillance scenes.Because the low-resolution photos taken in surveillance scenes have a wider distribution in pose,illumination,and resolution,this paper proposes an improved super-resolution network,introducing the UNET with an automatic encoder structure with a skip connection layer.It can extract the image information on multiple scales.The experiment proves that the improved super-resolution network can restore the face images taken from the surveillance scenes well,and can achieve high recognition accuracy in the face recognition task under the surveillance scenes.