Research On Super-resolution Of Face Images Based On Dual-branch

Face image as an important carrier of face features,its resolution reflects the richness of details.Higher resolution means the outlines of the face is clearer,the characteristics are more obvious,and the more conducive to the identification of the individual.Therefore,the high-resolution images of face have very important applications in security and other fields.However,in real scenes,the quality of the obtained face images are often poor due to the limitations of the imaging equipment,environmental interference,and loss of data transmission.The existing face super-resolution methods have a good effect on simulated data,but their effect on actual images are not ideal.Therefore,the research on superresolution technology of face images has important practical value and theoretical significance.For single frame face image and practical application scenario,in this paper,the superresolution reconstruction method based on deep learning is studied from two aspects of effective feature extraction and network construction.Firstly,in order to extract features better,a new residual attention block is designed in this paper.The block analyzes channel information of input features to guide the extraction of residual features.Meanwhile,in order to constrain the structure of reconstructed face images,this paper introduces the feature loss of the face images in the training of the reconstruction network.Based on the proposed residual attention block and the feature loss of the face images,we proposed an end-to-end super-resolution network for face images.Secondly,aiming at the problems that the existing algorithms perform well on data with known degradation methods,but the reconstruction quality of the face images in the actual field are poor,this paper combines the end-to-end super-resolution network proposed and designs a dual-branch super-resolution network structure.This structure builds a simulated degradation network to simulate the distribution of face images in real scene.Then simulated images are sent to two branches of the network to complete the constraints on the simulated image structure and the recovery of high-frequency information.Finally,in order to verify the effectiveness of the algorithm proposed in this paper,we use an open source database for testing.The experimental results show that the dual-branch network proposed in this paper can reconstruct more details of the face,and ensuring the authenticity of the reconstructed face image.