Research On Face Super-resolution Techniques In Real Scenes

In the field of computer vision,face analysis has always been a mainstream research topic.However,due to the constraints of the lighting conditions,target motion and the sensor itself,only low resolution face images can be collected.These images are lack of face details,so it is difficult to carry out further face analysis.Face super-resolution reconstruction technology recovers high-resolution face images by extracting and utilizing the information in low-resolution face images,which can successfully enhance the application value of low-resolution face images.However,the existing face super-resolution models generally obtain low-resolution images by manually down-sampling,which can not construct an effective training sample set of super-resolution model.In practical testing,the performance of super-resolution model often degrades seriously due to the difference of blur kernel distribution in the process of degradation.Starting from deep learning the degradation process of face image,this paper proposes an two-stage network based on deep learning.Firstly,it learns the degradation process of real face image by a generative adversarial network,and then constructs an effective training sample set and send it to super-resolution generative adversarial network to reconstruct high-resolution face image,which can effectively improve the quality of reconstructed face image.The face super-resolution network proposed in this paper consists of two confrontation subnetworks.One is face simulation degenerate sub network: a degenerate transformation network is trained by using high-resolution face images and real low-resolution face images to simulate the degradation process in real scenes to generate more real low-resolution face images.The degenerate transform adversarial sub network proposed in this paper considers blur kernel,noise and face landmark information at the same time to maintain the consistency of high and low resolution face image geometry,and also uses dense connection structure to enhance feature reuse in the network.The other is super-resolution reconstruction sub network:the low-resolution image generated by face simulation degenerate sub network and the corresponding high-resolution image reconstruct the final super-resolution face image through the super-resolution reconstruction adversarial sub network.The reconstruction network uses the super-resolution module with attention component to further improve the quality of generated face image.Qualitative and quantitative experiments were performed on SFLRS from LS3D-W and RFLRS from Widerface between our model and other methods.Quantitative experiment: our model achieves PSNR accuracy of 21.68 based on synthetic data set SFLRS and FID accuracy of 12.53 based on real data set RFLRS.Qualitative experiment: this model recovers more face details and the visual effect is more clear and real,which further shows that the performance of this model is still good for the data set in the real scene.Furthermore,the ablation studies valid the effectiveness of the proposed components.